AI Is Becoming Part of the Retail Investing Workflow

The article interviews several retail investors and shares their experiences of both making and losing money, which keeps it objective rather than one-sided. It also highlights a range of professional opinions on the trend, from bringing liquidity to previously overlooked asset classes to more cautious takes on how the lack of risk controls and oversight could pose real problems for retail investors, since these models are not built to withstand volatile environments. One investor quoted in the piece described using AI less as a stock picker and more as a discipline tool: “It removes all my biases, it removes all my fear, it removes all the emotions that a human would bring into a trade.” Others quoted were more careful, letting AI do the research work but still leaving the actual trading decisions to humans, or simply not allocating a large percentage of their portfolio to AI-assisted investments.

Surprisingly, I found the article a bit underwhelming. I was hoping for more insight into how this trend is actually changing market dynamics.

AI May Be More Useful for Discipline Than Stock Picking

To dig a bit deeper, I pulled a few more articles on the topic. First, I came across the Business Insider piece “Fear, greed, and Claude: Meet the day traders vibe coding AI bots to beat the market.” Similar idea, with multiple retail traders interviewed and sharing their experiences. All of them arrived at a common conclusion: an investor needs prior knowledge of the financial markets before letting a bot invest for them.

As Brendan Li, a former banker who now runs a vibe-coding community for traders, put it: “AI tools can’t help a trader who doesn’t understand the fundamentals of the market,” which is why he avoids working with beginners. “AI can’t give a trader a winning strategy,” he added. Instead, the tool tends to help traders who already have a strategy but struggle to stick to it, or who want to test a few variations of one.

Dr. Reid Daitzman, a psychologist who has been trading for 50 years and built his own signal generator with ChatGPT, made a similar point: building the program was not difficult, but it does require the investor to be fluent in the markets for the bot to actually be useful.

That lines up with what I think is the real story here: it is less about AI finding opportunities a human would have missed, and more about the emotional side of investing.

Daitzman coaches traders alongside his psychology practice, and he knows the emotional mistakes investors make firsthand. Traders typically come to him in moments of despair, often after losing large sums, and their issues usually trace back to not handling their emotions: revenge trading, trying to make up for a loss with increasingly aggressive trades, overtrading, undertrading, and the like.

“It helps control fear and greed,” he said of why the program worked for him. “If you have a buffer between you and the screen, it gives you perhaps a sense of security.”

René Balke, another trader who now runs fully automated accounts, echoed this. Some of his results are down to the broader run-up in the major indexes, but much of the benefit has simply been following his own strategy without falling victim to his own emotions.

Using AI seems to help eliminate irrational investment behavior more than it helps find better trades.

Retail Adoption Is Growing Quickly

The adoption numbers back this up. According to eToro’s Retail Investor Beat survey from late 2025, a survey of 1,000 American retail investors, 58% are now using AI tools to help build their portfolios, and 30% use AI to pick or alter investments outright, up 75% in a single year. When asked why they are open to letting AI weigh in on their investments, the top reasons were saving time on research (48%), a belief that AI is the future of investing (46%), a belief that AI will make better decisions than they would (38%), a belief that it will pick better investments than a fund manager (30%), and cost relative to a fund manager (25%).

Separately, Investing.com surveyed 938 US-based retail investors and found that 62% are already using AI tools to inform their investment decisions, with 53.5% using them specifically for investing-related research, including 30.6% using AI-powered research or stock-screening tools and 62.4% using AI to research stocks or other assets.

The Risk of Vibe-Coded Investment Tools

With so many articles highlighting this trend, I am particularly concerned about non-technical professionals building investment tools through vibe coding.

As I mentioned earlier, a well-designed investment model needs built-in risk controls and oversight. Without understanding the code, how do investors make sure those controls actually exist and can withstand volatile conditions?

As a quant, I use Claude Code often in my own workflow, but I still prefer to review the code myself and spend the time to understand what it is actually doing. Vibe coding can be great for getting something up and running in the moment, but it is far from production-grade code that stands the test of time.

A friend of mine once told me she dreads it when a colleague hands her a vibe-coded solution and simply says, “It works.” The code is rarely production grade. A standalone piece of code working in isolation does not mean it will still work once it is integrated into existing infrastructure. There are too many things that can break.

How to Use AI Without Blindly Trusting It

There are a few things non-technical professionals can do to make sure their AI-produced code holds up.

First, build version control into the coding workflow. Clearly document every change and every iteration, and preserve the ability to roll back to a previous version.

For investors specifically, start with paper trading and allocate only a small amount of capital at first. Do not simply trust the strategy an AI suggests. Make sure you understand it inside out before putting it into practice.

AI can be the brain behind the strategy, but you still need to be a good student and learn how it actually works.

Lastly, act as your own QA. Build out test cases to check the code’s quality, not just for normal market conditions but for extreme scenarios too. Also build test cases around infrastructure issues, like a power outage or the market data feed going down.

You do not need to be able to read the code, but you do need to understand how your AI tool will behave when things go wrong.

The Bigger Opportunity for AI in Retail Investing

Anyway, I am genuinely excited to see so many AI applications emerging in the retail investing space. What does your investment workflow look like now? Have you incorporated any AI into it?

First, I came across a Bridgewater engineering talk at LangChain’s Interrupt Conference and found it fascinating. I often get asked about real, practical AI applications in the investment industry. Bridgewater’s case demonstrates that AI can meaningfully improve investment workflows. In this particular talk, the focus is on deep exploratory research.

In addition, inspired by research published by DeepVest and Daloopa, I decided to learn more about the various benchmarks evaluating frontier models on finance-related tasks.

What Bridgewater actually built

In a talk at LangChain’s Interrupt Conference, Brendan McManus, Santi Weight, and Michael Ran walked through the architecture of PAT (Pocket Analyst Tool). It’s worth understanding because it’s not simply “we plugged in Claude and called it a day.”

Here are the takeaways.

The foundation is a well-structured internal database that combines both structured and unstructured data, built before any agent ever touches it. PAT’s agents then use the same tools Bridgewater’s own investors use internally. The agent isn’t reasoning against a stripped-down toy version of the firm’s research stack. It has the same context as a human analyst.

A few design choices stood out to me. There’s a feedback loop built directly into the product. When something is off, users can flag it through a “Teach” button, and that correction feeds back into how the system behaves going forward. Permissions are also enforced at the tool level, not just the user level, so an agent acting on an employee’s behalf cannot pull data that the employee shouldn’t have access to in the first place. That’s a real answer to a problem many AI deployments in finance haven’t actually solved.

The part I keep thinking about is the continuous learning loop. When something fails, PAT saves that failure state so the second attempt is faster, not because the model got smarter, but because the system remembers what didn’t work. Compare that to a frontier model in a fresh chat window, which starts from zero every time with no memory of the mistake it made yesterday.

On the coding side, Bridgewater reports that code generated by its coding agents is 95% reproducible, which is close to the level of determinism you’d want for anything supporting real portfolio decisions.

Bridgewater’s approach also got me thinking about a broader question: how well do today’s frontier models actually perform on finance-specific tasks? Around the same time, I noticed that two of my favorite MCP providers had published benchmark studies comparing frontier models on real investment workflows. That sent me down a rabbit hole looking for other finance-specific benchmarks and comparing what they found.

Two of my favorite MCP providers got me thinking

I noticed something this month: two MCP providers I follow closely, DeepVest and Daloopa, each published research benchmarking their own tools against frontier models.

DeepVest tested ChatGPT, Gemini, Perplexity, Grok, and Claude on ten real investment workflows, including drawdown analysis, backtesting, portfolio optimization, and macro stress testing. It found that frontier models failed roughly 85% of the workflows, mostly because of hallucinated numbers or an inability to compute correctly.

Daloopa ran a narrower test on financial data retrieval and found that plain ChatGPT achieved 63.8% accuracy, versus 94.2% for Claude connected through Daloopa’s MCP.

It’s worth pointing out that both companies sell products that benefit from those conclusions, so neither study is entirely neutral. Still, they prompted me to look for benchmarks with less of a stake in the outcome. That search led me to three more worth cross-referencing: Hebbia, Rogo’s Big Finance Bench, and Vals AI’s Finance Agent v2.

Hebbia went first. Its Financial Services Benchmark runs each question 50 times per model to ensure the results are statistically significant before drawing conclusions, a level of rigor that many vendor benchmarks don’t bother with. Across more than 600 real investment banking, private equity, credit, and public equity workflows, the results showed that no single model wins every task. Claude 4 Opus and GPT-4.1 perform best at extracting exact numbers from documents, while o3 and GPT-5 perform best at reasoning and summarization. Paying more also doesn’t consistently buy better output.

Rogo followed with Big Finance Bench, which includes 928 questions written and rubric-graded by 52 finance practitioners and reviewed by 12 senior panelists. It evaluates models in two ways: a rubric score, which gives partial credit for each correct step, such as finding the right filing, applying the correct definition, or performing the correct calculation, and a final-answer accuracy score, which only counts if the final answer is exactly correct.

The gap between those two metrics matters. Across all ten models tested, rubric scores averaged about 16 points higher than final-answer accuracy. In other words, a lot of otherwise correct reasoning is lost because of one mistake in the final step.

Then there’s Finance Agent v2 from Vals AI, an independent evaluator with no in-house model to sell. It includes 927 questions designed to test whether an LLM can perform the work of an entry-level financial analyst. Like Big Finance Bench, it uses rubric scoring, but it goes one step further. Agents are given a set of tools, and the benchmark also tracks how many tool calls each model requires to reach an answer, measuring both accuracy and efficiency.

Cross-referencing all five

Looking across all five benchmarks, a few themes emerge.

Every benchmark, whether vendor-run or independent, agrees that no single frontier model wins across every task type. Each model has different strengths depending on whether the task involves extraction, reasoning, or multi-step calculation.

Most of them also push back on the instinct to simply use the most expensive model. Rogo found that the same benchmark task costs roughly $1.26 using Claude Opus 4.7, versus about two cents using a cheaper model, while producing similar results once the model has the right documents and definitions available.

Hebbia reached a similar conclusion from another direction. GPT-4.1, a less expensive model, matched top-tier performance on several tasks.

Daloopa’s findings suggest the gap isn’t really about the model itself. It’s about what the model has access to. The same Claude model improved from 63.8% to 94.2% accuracy once it was connected to grounded financial data.

Both Big Finance Bench and Finance Agent v2 maintain live public leaderboards that are worth bookmarking as new models are released. Finance Agent v2 is particularly useful because it breaks results down by model, cost, and task category, making it easier to match the right model to a specific workflow rather than simply looking at an overall ranking.

Wrapping up for this week

Bridgewater built PAT around the idea that the model shouldn’t be trusted to freehand everything. Instead, it relies on a real internal database, existing investment tools, permissions enforced at the tool level, a feedback loop, and a memory of previous failures.

This week’s benchmarks, from both vendors and independent evaluators, reinforce why that approach makes sense. No single model wins across every task, cost doesn’t necessarily correlate with accuracy, grounded data matters more than model selection, and even the best systems become less reliable as tasks move from retrieval toward judgment.

What I’ll be watching next is whether more vendors begin publishing granular, task-specific leaderboards instead of a single overall score. That seems to be where the most useful signal actually lives.

That’s it for this week.

I was excited to share what I’ve learned using Claude, and the response was overwhelmingly positive. It’s clear there’s a growing interest among finance professionals in incorporating AI into their daily workflows. Surprisingly, one question came up in both workshops:

How do we minimize token costs?

Last week, a researcher at Quesma described burning through his entire Claude Max 5x plan in about 30 minutes. He kicked off /deep-research with one big open-ended question, walked away, and came back to no results. The run had launched 111 agents and queued 123 claims for verification, but only 25 of them were ever checked before he hit his limit. No report. No synthesis. Just a very short, very expensive research session.

I found the story both fascinating and validating. I’m currently researching how Claude’s usage limits actually work, so watching someone hit the wall in real time felt like a preview of the questions I was trying to answer myself.

This week I want to talk about tokens: what they are, what they cost, why your usage limits reset when they do, and the one habit that seems to separate people who run out of runway from those who don’t.

What a token actually costs you

Every time you send Claude a message, or Claude reads a file, calls a tool, or writes a response, that text is broken into tokens: small chunks of language the model processes. If you’re curious about how tokenization works, Tiktokenizer is a useful tool for visualizing how text is split into individual tokens. You are billed, or your usage allowance is reduced, based on how many tokens go in and how many come out.

Here’s the detail that trips people up: output tokens cost about 5x more than input tokens across every Claude model. Reading is cheap. Writing, reasoning, and generating answers are not. A model that reasons through a problem step by step, calls a dozen tools, and writes a long report is spending heavily on the expensive side of that ratio the entire time.

That is exactly what happened in the Quesma case. /deep-research doesn’t just answer a question. It plans, searches, verifies, and writes, all of which are token-intensive operations. It also does this across many agents at once. With 111 agents running in parallel, even a plan designed for heavy use will reach its limit quickly.

What Claude models actually cost (per million tokens)

Source: Anthropic pricing, current as of July 2026.

The spread here matters more than it looks. Fable costs 10x more than Haiku for both input and output tokens. If you’re running a task that Haiku could handle just as well, defaulting to Fable isn’t getting the best model for the job. It’s paying a 10x premium with little or no benefit.

How usage limits and resets actually work

This is the part I’ve been digging into over the past week. Here’s what I’ve learned so far.

Claude’s usage on paid plans works on two clocks at once.

A 5-hour rolling session window. This starts with your first message and covers short bursts of heavy use. If you’re deep into a research session or a large coding task, this is usually the first limit you’ll encounter.

A weekly cap. This resets at a fixed day and time tied to your account, not when you happen to start using Claude. You receive your full weekly allowance back on that schedule every cycle.

If you hit a limit, you have three options: wait for the reset, upgrade your plan, or purchase usage credits, which allow you to continue working at standard API rates instead of stopping completely.

None of these are wrong choices. But knowing which clock you’re up against, the 5-hour window or the weekly cap, changes how you plan larger tasks.

The fix: route by complexity, not by default

Here’s what the Quesma researcher changed, and it’s the part worth taking away.

He subscribes to several AI tools, giving him the flexibility to route tasks across different providers.

Instead of sending everything to the most capable, and most expensive, model by default, he assigned each model a specific role in the research process:

• Find: A fast, inexpensive model performs the search.

• Verify: A more accurate model checks what was found, never the same model that found it.

• Judge and plan: The most capable model is reserved for breaking down the problem and resolving disagreements.

• Deep research: Runs last, using findings that have already been verified instead of exploring blindly.

The result, in his words, was roughly 10x more research runtime for the same subscription cost. Not because he found a loophole. He started routing tasks by complexity, letting less expensive models handle routine work and saving the premium model for the jobs that truly required it.

The same principle, applied at bank scale

According to a report published by Klover.ai, Goldman Sachs doesn’t leave this kind of routing to individual judgment. It built the decision into its infrastructure.

The firm’s internal GS AI Platform is a centralized, multi-model gateway that gives employees access to a portfolio of models, including GPT, Gemini, Claude, and open-source models such as Llama, all behind a secure internal platform.

Rather than betting the firm on a single provider, the platform routes each task to whichever model is most effective and most cost-efficient for that specific job.

The reasoning is the same as the Quesma researcher’s, just at enterprise scale.

Goldman describes the approach in terms of three benefits:

• Capability diversification: Use the right model for the task instead of relying on one model for everything.

• Risk mitigation: Avoid becoming dependent on any single vendor’s pricing or performance.

• Future-proofing: Add new models without rebuilding the platform.

On top of that routing layer sit role-specific AI tools designed for real workflows rather than a generic chatbot.

The Banker Copilot helps with drafting and financial modeling for deal work, reportedly reducing M&A preparation time by around 40%.

For research and analysis, the firmwide GS AI Assistant supports document summarization and first-draft writing, with reported productivity gains of up to 25% in some use cases.

None of this works by sending every task to the most expensive model available. It works because the routing decision, using inexpensive models for search, more capable models for verification, and premium models only for judgment, happens automatically before anyone has to think about it.

You don’t need a billion-dollar AI budget, a dedicated engineering team, or 46,000 employees to apply the same principle.

You need to develop the habit of asking whether a task actually requires the most expensive model before you hit Send.

Five simple habits that save tokens

Routing is the macro fix. These are the micro ones: small habits at the keyboard that add up over a week of daily use.

1. Batch your questions

Every new message resends the entire conversation history as context. If you have three related questions, ask them in one prompt instead of three separate messages. You pay for that history once instead of three times.

2. Paste the excerpt, not the file

If you only need page four of a forty-page filing, paste page four. Input tokens scale with everything you include, whether the model needs it or not.

3. Set an output limit

Output tokens are the expensive side of the equation, costing roughly 5x more than input tokens. Asking for “three bullet points” or “under 200 words” limits the expensive part of the exchange before it even begins.

4. Keep related questions in the same conversation

Prompt caching can reduce input costs by up to 90% on content that stays the same, but only if it remains in the same conversation instead of being pasted into a new one each time.

5. Start a new conversation for a new task

While related questions belong in the same conversation, unrelated tasks do not. A single thread that wanders across multiple topics carries all of that history into every new message. Three short, focused conversations often cost less than one long, unfocused one.

The takeaway

Tokens are not an abstract billing detail. They are the unit of cost behind every AI-assisted research task, every model comparison, and every long-running agentic workflow.

Understanding the difference between input and output tokens, knowing what each model tier costs, and recognizing whether you’re working against a 5-hour usage window or a weekly limit will help you get much more out of the same subscription than simply defaulting to the most expensive model every time.

I’m still learning more about how Claude manages usage limits, and I’ll continue sharing what I discover. But one lesson is already clear: the cheapest way to use AI isn’t always to buy a cheaper plan. It’s to use the right model for the right task.

Editor’s Note

Thank you to everyone who has supported Mindful Data as a paid subscriber.

Over the past few months, my professional commitments have grown significantly. Rather than continue the paid subscription without being able to consistently create the additional content it deserves, I’ve decided to refund all subscription payments made to date and make every previously paid article publicly available.

Mindful Data isn’t going away. I’ll continue publishing free articles on the usual schedule, and all tutorials will remain openly accessible.

Thank you for your support, your trust, and your encouragement. I look forward to continuing this journey with you.

So here’s where things stand: the broker race to MCP has been impossible to ignore, there’s a security angle worth flagging, and I have a short list of things I want to test before share with you.

The Broker Race to MCP

The most interesting development this month is not a new data tool. It is execution.

ThinkMarkets launched an MCP server called ChelseaAI that lets traders connect any AI client to their trading platform and place trades without logging into the platform directly. cTrader shipped official MCP servers for AI-powered trading, and Webull took a different approach by rolling out its own MCP server that lets investors use plain-language instructions for market data queries and order entry, with the goal of expanding AI-powered quantitative tools to non-technical users.

Around the same time, IG Group and eToro announced or deployed their own MCP servers too. This is not one broker experimenting. It is the entire category moving at once.

What struck me is how much thought has gone into guardrails, and every company does it differently. ThinkMarkets built in circuit breakers and hard limits so AI cannot execute an entire account in a single trade, and traders can revoke execution permissions entirely while still allowing AI to analyze markets. AI also cannot deposit or withdraw funds on clients’ behalf.

IG Group, on the other hand, allows only read-only access through its MCP integration. No trade execution is permitted through its AI layer. That distinction, between an AI that can look and an AI that can act, is going to matter a lot as this space matures.

Not All MCP Access Looks the Same

Watching this race, I noticed the brokers are not all building for the same reason, and that changes how I think about evaluating any MCP server.

Deriv took a different approach than the others. Instead of gating its MCP-powered tools behind an existing account, it launched TradersView, a login-free market analysis platform that generates AI signals for anyone who lands on the page. No signup required. It attracted 20,000 active users and 33,000 page views in its first week.

This is not a retention feature for existing clients. It is an acquisition funnel, and a pretty effective-looking one.

Compare that to something like Quiver Quantitative, an alternative data provider, which gates its MCP server behind a paid API key. There, the data itself is the product, so there is no reason to give it away.

The lesson I am taking from this is that before I trust an MCP server, I ask myself whether it is free because the vendor wants me to try their data, or free because they want me to become a user of something else entirely. Both are reasonable business decisions. They just come with different incentives regarding what the tool is optimized to show me.

The Part I Am More Cautious About: Security

If you know me, you know I do not take AI output at face value, especially when important decisions are involved.

Microsoft’s security team recently flagged that MCP tool descriptions can be manipulated to steer an agent into leaking data or taking unintended actions. That risk becomes much more serious when the agent has standing access to a trading or account system.

One analysis I read put it well: a single prompt injection or stolen session can turn into a reportable financial event before a human is ever in the loop.

Microsoft illustrates this with a finance operations example. A third-party MCP server keeps the same name and summary, but quietly changes the tool description to instruct the agent to retrieve unpaid invoice data and attach it to an enrichment request, all without the user noticing anything unusual.

Here’s the action plan, based on Microsoft’s recommendations.

Before approving any agent that connects to business systems through an MCP server, start with inventory. Identify every approved publisher and server in use, turn off broad “allow all” connections, and scope each agent to only the specific tools it actually needs, nothing broader by default.

Once a server is approved, baseline it. Record the tool descriptions, schemas, and permission sets at the time of deployment, since these are exactly what can be quietly altered later without anyone noticing. Any change to that baseline should trigger a review before the modified tool touches a sensitive workflow, not afterward.

From there, the focus shifts to monitoring the action path itself. Apply data loss prevention policies to tool-call parameters, require human approval for high-impact actions, give agents their own non-human identities rather than reusing human credentials, and correlate telemetry between servers and agent behavior so unusual patterns actually surface. That same scrutiny needs to extend beyond MCP servers to adjacent surfaces, like AI-branded browser extensions, which have already been used to abuse trusted software channels.

Treat tool annotations as a hint, not a control. The MCP specification itself says clients must treat annotations as untrusted unless they come from a trusted server. A “read-only” label is not a substitute for real access controls, sandboxing, or an approval workflow. It is simply a label that a malicious server could set to whatever it wants.

None of this is a one-time gate. As agents gain the ability to read, write, send, and modify business data, approval has to become a continuous process: knowing at any moment which tools are connected, what their descriptions currently say, what data they can reach, and which of their actions still require a human to sign off.

A Few Things on My List

Given how behind I am on actually testing these tools rather than just reading about them, here are the two I want to carve out time for.

Deriv. Given the traction TradersView received, I want to see what the underlying MCP server offers once you connect an actual account, not just the login-free demo.

Quiver Quantitative. Congressional trades, insider activity, government contracts, lobbying spending, dark-pool short volume, and 13F changes, all in one server. This also happens to be a good teaching example because its official setup does not use a simple URL connection. Instead, it requires a small local bridge to pass the API key as a header, which is a connection pattern I have not needed for any other tool so far.

I’ll report back on both once I’ve actually tested them, hopefully before next week.

Final Thoughts

The theme I keep coming back to is that MCP is moving from an “interesting protocol” to actual financial infrastructure. And infrastructure needs both access and guardrails to be useful.

The reason is simple. Unlike software developers, finance professionals depend on external data every day: market data, company filings, earnings transcripts, research reports, and internal documents. AI becomes much more useful when it can access those resources directly instead of relying only on its built-in knowledge.

Rather than debating whether MCP still matters, I wanted to see what it could actually do.

So I connected two finance-focused MCP connectors, Deepvest and Daloopa, to Claude Desktop and tested them using Microsoft as the anchor company. I wasn’t looking for a winner. I simply wanted to understand how each connector works, what problems it solves, and where its limitations are.

What I found surprised me. Although both are finance MCP connectors, they take completely different approaches. One delivers analyst-ready insights with a single prompt. The other provides source-linked financial data that you can trace directly back to the original filing.

Here’s what happened.

What I Tested

I connected both connectors to Claude Desktop and tested them across several common investment research tasks:

• Earnings analysis

• Technical analysis

• Asset analysis

• Macroeconomic analysis

• Financial statement analysis

• Cross-asset analysis

Both connectors were live, and all observations below come directly from testing their MCP implementations.

Deepvest: It Does the Analysis for You

I mentioned Deepvest in a previous post because I was impressed by its AI agents for portfolio optimization and backtesting. Through MCP, many of these capabilities are now available directly inside Claude.

The overall experience is straightforward. You ask a question, and Deepvest returns a structured answer. Instead of exposing raw datasets, it performs much of the analysis before presenting the results.

Here are the features I tested. The MCP connectors provided the financial data and analytics, while Claude organized the output into a more readable format, including summaries and visualizations where appropriate.

Earnings Analysis

I asked Deepvest to analyze Microsoft’s latest earnings.

Instead of simply retrieving the earnings release, it returned a structured report that included earnings performance versus expectations, key financial metrics, management commentary, and the major themes discussed during the earnings call. Reading the output felt more like reviewing an analyst’s earnings note than searching through company filings.

When I’d use it: Quickly understanding an earnings release before reading the full transcript.

Technical Analysis

Next, I asked Deepvest to perform technical analysis on Microsoft.

The output combined multiple trend and momentum indicators into a single summary while highlighting important support and resistance levels. Rather than forcing me to interpret dozens of individual indicators, it distilled everything into an overall technical assessment.

When I’d use it: Getting a quick technical snapshot before digging deeper into price action.

Asset Analysis

The asset analysis feature generated a concise risk and performance profile for Microsoft.

It summarized return statistics, volatility measures, drawdown metrics, benchmark comparisons, and other portfolio analytics in one organized table. It looked very similar to the type of summary a quantitative analyst might prepare before a portfolio review.

When I’d use it: Reviewing an individual holding before an investment meeting.

Macroeconomic Analysis

I then asked Deepvest to summarize the current macroeconomic environment.

The response included both a visualization and a written summary. It combined inflation, interest rates, GDP, credit spreads, and consumer sentiment into a single dashboard before summarizing the overall macro backdrop and the key economic themes driving markets.

This is the type of dashboard many analysts build manually in Excel. Deepvest assembled it from a single prompt.

When I’d use it: Preparing for a macro discussion or investment committee meeting.

Financial Statement Analysis

Deepvest also includes a financial statement analysis feature.

Rather than pulling raw SEC filings, it returned a summarized view of Microsoft’s financial statements, including profitability, margins, leverage, and other key financial metrics in a clean, readable format.

The emphasis is on helping you understand the business quickly rather than verifying every number against the original filing.

When I’d use it: Getting a fast overview before diving into the financial statements.

Cross-Asset Analysis

This was probably my favorite feature.

I asked Deepvest to analyze the relationship between macroeconomic variables and the performance of Microsoft, QQQ, and SPY.

Instead of simply returning historical data, it performed the statistical analysis, identified the macro factors most closely associated with asset performance, and summarized the results in plain English.

This is the kind of analysis that would normally require gathering multiple datasets, writing code, and running regressions yourself.

When I’d use it: Exploring macro relationships and generating new research ideas.

One Limitation

I also tested the portfolio optimization feature because it was one of the capabilities I was most excited about.

Unfortunately, the returned portfolio allocated equal weights across every position rather than generating an optimized allocation. That wasn’t the behavior I expected, so I removed it from my workflow for now.

Daloopa: It Gives You the Data

Daloopa takes a very different approach from Deepvest.

Instead of generating analyst-ready insights, Daloopa focuses on providing high-quality, standardized financial data that you can trust. Its platform organizes company financials into structured data sheets, making it easy to explore historical financials without manually extracting information from SEC filings.

What makes the platform stand out is its traceability. Every reported value is hyperlinked to the exact location in the original filing, allowing you to verify a number with a single click. Rather than asking you to trust the output, Daloopa lets you see exactly where every figure comes from.

The MCP connector brings this experience directly into Claude. You can discover companies, retrieve standardized financial data, and access SEC filings without leaving your AI workflow, while still maintaining a direct link back to the original source.

Think of Deepvest as an AI research analyst that delivers answers.

Think of Daloopa as a trusted financial database that gives you the evidence behind those answers.

The Workflow

Using Daloopa requires a few more steps.

First, you identify the company.

Next, you discover the financial series you’re interested in.

Finally, you retrieve the corresponding financial data.

Once you’ve identified the series you use regularly, subsequent queries become much faster. The extra setup is the price you pay for accessing structured, source-linked financial data.

Financial Data Retrieval

Once the company and financial series are identified, Daloopa returns standardized financial statement data directly from SEC filings.

Unlike Deepvest, the emphasis isn’t on summarization. Instead, every figure is linked back to its original source, making it easy to verify any number with a single click.

For analysts producing client reports or investment memos, that level of traceability is extremely valuable.

When I’d use it: Building financial models or validating numbers before publishing research.

Document Discovery

I also tested Daloopa’s document discovery feature.

With a single query, it returned Microsoft's recent 10-Qs, 10-Ks, earnings presentations, earnings call transcripts, and other SEC filings. Each document includes a document ID that can be passed to the get_document_content tool to retrieve the full filing or transcript directly within Claude.

Instead of searching EDGAR manually, the documents become immediately accessible inside your AI workflow.

When I’d use it: Finding source documents without leaving Claude.

One Limitation

Daloopa currently works one company at a time.

If you’re analyzing a portfolio with multiple holdings, you’ll need to repeat the workflow for each company. Some datasets also require higher subscription tiers, so it’s worth confirming your access before building automated workflows around them.

The Structural Difference

After testing both connectors, I realized I wasn’t comparing two MCP connectors.

I was comparing two different philosophies.

Deepvest performs the analysis and gives you the answer.

Daloopa gives you trusted data so you can perform the analysis yourself.

Neither approach is better. They simply solve different problems.

My Takeaway

If I need to understand a company quickly, I’d start with Deepvest.

If I’m preparing research that will be shared with clients or investment committees, I’d reach for Daloopa.

In practice, I think they complement each other remarkably well.

More importantly, this experiment reinforced why I still believe MCP matters for finance.

The protocol itself isn’t the story.

The real value is making specialized financial data and research tools available directly inside AI workflows. Whether you’re retrieving SEC filings, running macro analysis, or validating financial statements, MCP allows AI to work with the same information that investment professionals use every day.

And that’s where I think the biggest opportunity lies.

What’s Next

In my next paid post, I’ll show how I combined Daloopa and Financial Modeling Prep with Claude to analyze an entire client portfolio.

The workflow includes approval checkpoints at every stage, automatically identifies the portfolio’s biggest contributor, retrieves company fundamentals, analyst price targets, insider trading activity, and generates a client-ready summary dashboard.

I’ll also share the complete prompt, walk through the workflow step by step, and include the final HTML report so you can build and adapt it.

→ Continue to the paid post: [Building a Client Portfolio Analysis Workflow with Claude, FMP, and Daloopa]

One capability I believe is still missing is accessible and professional-grade investment analytics built directly into MCP. Imagine being able to run a backtest or portfolio optimization entirely within your chat window, without switching to another application. As more online traffic shifts toward AI assistants, I can envision people relying on AI platforms for many tasks instead of navigating multiple websites and applications, including making financial decisions.

This is not a polished demo. It is simply an experiment, and there is still plenty of room for improvement. I still need to refine both the prompts and the outputs. With the rise of loop engineering, even prompt generation may eventually become automated, and that is one of the research directions I plan to explore.

Step 1. What FMP and Daloopa actually do in this workflow

Before we get into the prompt, it helps to understand what each connector does and why I chose it.

Financial Modeling Prep (FMP)

Financial Modeling Prep (FMP) provides a broad set of market data APIs covering historical prices, technical indicators, analyst estimates, insider transactions, financial statements, and news. I am currently using the free trial, which has limited access, but it is more than sufficient for building this workflow and experimenting with portfolio analysis.

The workflow uses five FMP endpoints:

chart — Retrieves daily adjusted closing prices over a specified date range. This serves as the foundation for all return and risk calculations, including cumulative return, annualized return, volatility, Sharpe ratio, and maximum drawdown.

technicalIndicators — Retrieves RSI, 50-day SMA, 200-day SMA, and ADX for each ticker. Claude combines these indicators to classify each holding as bullish, neutral, or bearish. Since FMP returns one indicator per request, the prompt handles the aggregation automatically.

analyst — Retrieves the latest analyst ratings and price targets by firm. The workflow summarizes the ten most recent ratings, calculates the consensus recommendation, computes the average price target, and estimates the implied upside or downside relative to the current price.

insiderTrades — Retrieves the ten most recent insider transactions for a given ticker, including the buyer or seller, transaction size, price, and value. The workflow also summarizes insider buying and selling activity over the previous 90 days to provide a simple sentiment indicator.

news — Retrieves five to seven recent news articles together with sentiment tags. Claude adds a brief summary explaining the potential relevance of each article to the company.

Daloopa

Daloopa specializes in structured fundamental data extracted directly from company filings. One feature I particularly like is that every data point includes a hyperlink back to its original source, making the analysis fully traceable and easy to verify.

In my previous free post, I explained how Daloopa’s MCP is structured and the multi-step process required to retrieve company data. This workflow builds directly on that foundation.

The workflow uses three Daloopa endpoints:

discover_companies — Looks up a company by ticker and returns the Daloopa company ID required for subsequent requests.

discover_company_series — Lists all available financial data series for the company, including total revenue, segment revenue, gross margin, EPS, free cash flow, and many others. This allows Claude to identify which metrics are available before requesting the data.

get_company_fundamentals — Retrieves up to eight quarters of the selected financial metrics. Every data point includes a source link back to the original SEC filing or earnings release, making the output auditable rather than just another number in a table.

Because Daloopa requires separate queries for each company, this workflow limits the fundamental analysis to the portfolio’s highest-impact holding, which is identified automatically in Step 4. This keeps the workflow efficient while avoiding unnecessary deep analysis of every position.

To make this portfolio analysis more reliable, the next step will be to build a data validation agent that verifies the numbers before they appear in the report. I also envision a personalization agent that adapts the writing style, examples, and vocabulary to match each client’s preferences. AI has the potential to make personalized financial communication accessible to everyone, not just wealthy clients.

Step 2. How I built the prompt

The final prompt you’ll see below is not how it started. Keep in mind that this is still an experiment. I simply looked at the financial data available through the MCP connectors and built the portfolio analysis around those capabilities.

The first step was straightforward. I told Claude I wanted to build a portfolio analysis tool for financial advisors using Daloopa and FMP. I then reviewed the available features from both connectors and decided what the workflow should include. I ended up building a step-by-step process that starts by reading the portfolio, pulls data from the MCP connectors, calculates portfolio metrics, and produces a client summary.

Next, I asked Claude to add a step-by-step approval loop. Instead of running the entire workflow at once, Claude pauses after each step and waits for your input. You can type ok to continue, provide feedback to make adjustments, skip to move on, or rerun to retry a step. That single change made the workflow much easier to debug because you can review each stage before it feeds into the next.

The second iteration focused on presentation. A table of numbers inside a chat window is useful for review, but it is not something you would hand to a client. I added an HTML artifact with tabbed navigation so each stage populates its own tab, along with Chart.js visualizations that are generated as the data is processed.

I then added an impact analysis to identify the stock that contributed the most to the portfolio’s performance. For that company, Claude combines Daloopa’s fundamental data with FMP’s analyst ratings and insider trading data to build a narrative explaining why that holding mattered and what may have driven its performance.

One final observation: a prompt this structured is only a short step away from becoming a reusable skill. Once you’ve tested it a few times and are satisfied with the results, you can package it as a Claude Skill so anyone on your team can run the same analysis with a single command, without needing to understand the underlying prompt.

Step 3. What each step does and what gets built

The workflow runs in seven steps. Here’s what each one does and what appears in the HTML artifact.

Step 1 – Asset allocation overview

No connector is needed. Claude reads the portfolio weights you provide and builds a holdings table with the ticker, company name, weight, sub-sector, and market cap category. The artifact displays a doughnut chart by ticker and a horizontal bar chart showing sub-sector concentration. This provides the foundation for everything that follows.

Step 2 – Portfolio performance and position-level risk

Using FMP’s chart endpoint, Claude retrieves three years of daily price history for all holdings. It calculates portfolio-level metrics, including cumulative return, annualized return, volatility, Sharpe ratio, and maximum drawdown, for both the 1-year and 3-year periods. It then breaks these metrics down by individual holding. The artifact includes a grouped bar chart of the portfolio metrics and a scatter plot showing each holding’s return versus volatility, with bubble size proportional to portfolio weight.

Below is a screen recording showing how Claude builds the Step 2 tab in real time.

Step 3 – Technical signals

Claude calls FMP’s technicalIndicators endpoint to retrieve the RSI, 50-day SMA, 200-day SMA, and ADX for each holding. Based on predefined thresholds, it derives an overall technical signal for every ticker and presents the results as a color-coded heatmap: green for bullish, yellow for neutral, and red for bearish.

Step 4 – Identify the top impact stock

No connector is required. Claude multiplies each holding’s portfolio weight by its 1-year return to calculate its weighted contribution, ranks all positions, and identifies the stock with the greatest impact on portfolio performance. That ticker is then passed automatically to Steps 5 and 6. The artifact displays a horizontal bar chart with the top contributor highlighted in gold.

Step 5 – Deep dive on the top stock

This step consists of three sub-steps. First, Claude retrieves eight quarters of total revenue, key segment revenue, gross margin, non-GAAP EPS, and free cash flow from Daloopa, with source links for every data point. Next, it pulls the ten most recent analyst ratings and price targets from FMP. Finally, it retrieves the ten most recent insider transactions. The artifact includes a revenue and margin combination chart, an analyst price target chart with a reference line for the current stock price, and an insider trading timeline.

Step 6 – News and context

Claude retrieves five to seven recent news articles from FMP, along with their sentiment tags. It also creates a combination chart that overlays the stock price with the segment revenue trend from Step 5, allowing you to compare business performance with market performance. Finally, Claude writes a short narrative explaining why this stock had the greatest impact on the portfolio, what the underlying business data suggests, and whether the current portfolio weight still appears justified.

Step 7 – Client report and summary dashboard

No connector is required. Claude synthesizes the results into a four-section client report covering portfolio performance, position highlights, risk profile, and outlook. Below the narrative, a 2×2 dashboard brings together the key charts from each stage into a single view. A download button allows you to save the complete report as an HTML file.

The prompt

Note:
The Excel version of this workflow uses the same prompt logic but outputs to a 7-sheet workbook through Claude in Excel. It works well, but it is significantly more token-intensive, so I am not covering it here.

If you run this workflow with your own client portfolio, I would love to hear what you changed or improved. Reply to this post or leave a comment.

What Comes Next

This workflow still requires you to stay in the loop. You paste the prompt, type ok after each step, and intervene if something goes sideways. That is intentional for a first build. Understanding what each step does and where it can fail makes you a better operator of the tool.

But the manual approval loop is not the destination.

The next evolution is loop engineering: structuring the prompt so Claude evaluates its own output at each step, decides whether to proceed or retry, and completes the full analysis without waiting for user approval. No more typing ok seven times. You paste the prompt, come back a few minutes later, and the report is finished.

Beyond that, one of my current research interests is building an AI-augmented quantitative research workflow. Portfolio analysis is only one layer. What interests me more is the entire research process: screening investment ideas, stress-testing a thesis against fundamentals and macroeconomic data, validating the results, and producing the kind of structured analysis that would normally take an analyst an entire day.

The connectors already exist. The challenge is designing workflows that combine them intelligently, add validation where it matters, and automate repetitive tasks while keeping humans focused on higher-value decisions.

This portfolio analysis is one small experiment toward that broader vision, and I expect many more experiments to come.

On June 8, Google rebuilt NotebookLM on Gemini 3.5 and gave every notebook a secure cloud computer that can write and run code to help users perform deeper research. The system now includes more than 100 curated software skills. New output formats were also introduced, including PDFs with charts and tables, Excel spreadsheets, and even CSV or JSON files. The research workflow has been improved as well. Users can start with loose ideas and questions in chat, and NotebookLM will guide them in building source libraries from there.

Unfortunately, access is still limited, so I have not had a chance to try it myself. I rely heavily on NotebookLM to fact-check a wide range of resources for my research, and I am looking forward to experimenting with the new features.

On June 9, Anthropic released Fable 5, the most powerful model it has made available to the public, and described it as the strongest finance model it has tested. The internet has already gone crazy over it, and there are plenty of impressive examples showcasing what Fable 5 can do.

When a week looks like that, the natural reaction is to assume the smartest model will now do your work for you. However, that is not true. Being AI-proficient was never about having access to the best model. It is about how you use it.

A new term, “model routing,” has emerged recently. Echoing my previous post about Goldman Sachs’ AI strategy of using multiple LLMs, model routing resembles a portfolio optimization problem. Depending on the use case, tasks are routed to different models to maximize performance, minimize token costs, or improve efficiency.

A Test, Not a Benchmark

The benchmarks for Fable 5 are impressive, but they are also Anthropic’s own. A lab score tells you what a model can do, not how it performs on your work.

So I ran my own test.

I gave Fable 5 and the prior flagship, Claude Opus 4.8, the same assignment:

• Summarize a quantitative research paper for a portfolio manager

• Reproduce its results table

• Audit the paper’s GitHub repository against its claims

• Plan a replication using a different set of financial securities in a different market

Same prompts. Same paper. Side by side.

The results were split, and that is the interesting part.

Fable 5 was much faster across all tasks and asked far fewer questions during the process. It reproduced the results tables more quickly and cleanly, and it produced a more detailed replication plan.

Opus caught more.

For example, it identified a discrepancy between the paper’s results table and the GitHub repository. Opus flagged it. Fable did not. I only knew Opus was right because I verified the claim against the live repository myself.

That is the whole lesson in one example.

The more polished answer was not the more reliable one, and the factor that settled it was a human verification step. Speed and fluency are not the same as accuracy, and neither model is trustworthy without someone who knows how to validate the output.

Two Talks, One Message

The same theme has surfaced repeatedly over the past few weeks. In fact, I heard it at two very different events.

First, I attended Snowflake Dev Day. I have attended for several years, and this year felt very different.

Over the past few years, attendees were excited about how AI would change the future. This year, the mood was noticeably different. With many attendees actively seeking jobs, there was a sense of vulnerability around AI. Some people no longer know how to approach the job search process. As both employers and candidates increasingly use AI, it seems to be taking longer to fill roles.

Despite that, the keynote talks were inspiring.

For the past few years, Andrew Ng took center stage and accurately predicted many of the most transformative AI innovations. This year, Peter Steinberger, creator of OpenClaw, took the stage and shared his perspective.

His message about AI agents was straightforward: they are tools, not people. They amplify your work, but they still require your expertise.

From other speakers, I heard a similar recommendation. Start with boring, difficult problems, or begin with functions that do not fundamentally change existing processes, such as research. Most people are obsessed with using AI to write code and build products. It may be time to think about how AI can support other parts of the software development lifecycle as well.

At a CFA Society Seattle event, Gavriella Schuster, Board Director and former Microsoft CVP, shared her perspective on transformation in the age of AI.

A similar idea emerged, but this time the framing was about careers.

Instead of thinking about your job as a single role, think about it as a portfolio of tasks. Some of those tasks, perhaps 30% to 50%, can be delegated to AI. The rest, including judgment, decision-making, empathy, and influence, remains yours.

Everyone should begin identifying the parts of their work that can be outsourced to AI, such as repetitive tasks, administrative work, scheduling, and coordination. Then shift your mindset from doer to designer, manager, or orchestrator.

By delegating routine tasks to AI, you create more time for meaningful work.

As machines take over more tasks, human connection becomes even more important. For example, teachers can use AI to assist with content creation and grading, allowing them to spend more time interacting with students.

Put those two perspectives together, and you get a practical playbook.

How to Actually Become AI-Proficient

Start with an audit.

Write down everything you do during a typical week and divide those activities into two categories: AI-eligible and human-required.

Then begin delegating the AI-eligible work to appropriate tools. For example, use NotebookLM for research and AI meeting assistants to help organize your day.

From there, the shift is from doer to designer.

Stop being the person who moves information around and become the person who decides what to do with it. Move from task executor to strategist. From coordinator to leader.

AI handles execution.

You handle direction.

The Future of Work Is Human

None of this is about competing with AI.

AI amplifies your value. It does not define it.

The professionals who pull ahead will not be the ones with access to the best model. They will be the ones who audited their work, redesigned their roles around what only humans can do, and kept their own judgment in the loop.

A flagship model and a free research tool both launched last week. Both are useful. Neither will tell you which tasks to delegate or catch the mistake that truly matters in the output.

That part still belongs to you.

And that is exactly the point.

So this week, pick one task. Audit it. Hand the repetitive part to a tool. Then verify the output yourself.

That is where your future role begins.

In this post, we will discuss how to turn a skill into something that runs on its own.

We have already built the economic-calendar skill. Now, we will learn how to schedule a task and make it run at a set time or frequency for you. In our economic-calendar example, we will create a scheduled task that runs the skill at 6:30 a.m. every weekday. When you wake up in the morning, the day’s economic calendar will already be waiting for you.

What a Scheduled Task Actually Is

A scheduled task is a standing instruction. You tell Claude what to do and when, and it runs without you having to start the conversation.

In the past, if you were not technical, you would typically need developers to set up recurring tasks for you. Most of the time, this required creating a cron job or using scheduling services such as AWS Lambda. With Claude, you can schedule tasks yourself, which can be a significant time-saver. As anyone who works in a corporate environment knows, getting additional resources often takes time.

There are two types of scheduled tasks, and the distinction matters:

Recurring tasks run on a cadence: every weekday morning, every Monday, every hour. The economic-calendar skill is a good candidate for this. The calendar changes meaningfully throughout the week, so a weekday morning run makes sense.

One-off tasks run once at a specific future time: “remind me in an hour” or “check this tomorrow at 3 p.m.” These are less interesting for our purposes, but it is worth understanding the difference because the setup conversation is different. A recurring task needs to consider cadence and what happens if something goes wrong on run 47. A one-off task does not.

Two Ways to Schedule a Task

Here is how you access scheduled tasks:

Cowork sidebar → Scheduled → New Task

Under New Task, there are two ways to schedule a task:

Create with Claude

You schedule a task by starting a conversation with Claude. Simply tell Claude what needs to be scheduled, along with the time and frequency, and it will take care of the rest.

Set Up Manually

A new task window will appear. You will need to name the task, provide a description, and tell Claude what it should do.

There are also additional configurations you can choose from:

Ask
Tell Claude whether it needs your approval before taking actions.

Default Model
You can choose which model the task should use. To maximize efficiency and minimize token usage, it is important to match the model to the complexity of the task.

For example, if the task is simple, Sonnet is usually sufficient. For more complex work, you may want to use Opus. The new Fable 5 performs well on long-running and complicated tasks, but it also costs roughly twice as much as Opus. At the time of writing, it is available only to Pro-tier subscribers and above. We do not yet know what will happen after June 22,2026, so I would not recommend setting up recurring tasks with Fable 5 until pricing and availability become clearer.

Frequency
Tell Claude how often the task should run and what time it should run. Claude also provides a gentle reminder to schedule tasks during off-peak hours whenever possible to help minimize costs.

Folder
Choose where Claude should access information and store outputs.

Turning the Economic-Calendar Skill Into a Scheduled Task

Now it is time to turn the economic-calendar skill into a scheduled task.

Here is roughly what the setup conversation looks like if you decide to create it with Claude:

“Schedule my economic-calendar skill to run every weekday morning at 6:30 a.m. ET. Save the output as a markdown file in my research folder so it’s there when I sit down.”

That is most of it. Claude handles the cadence, trigger, and execution. The skill itself, the one that searches MarketWatch, Investing.com, and Briefing.com and produces a grouped, chronological briefing, runs exactly as it did when you called it manually.

Here is what the setup looks like if you decide to configure it manually.

One deliberate choice here: I am keeping the output simple by saving it as a file that I can open when I sit down at my desk, rather than pushing it somewhere like Slack. Connectors deserve their own discussion, and a future post will cover what changes when the output goes somewhere other than a file.

For now, the point is simpler: the task runs, and the output is waiting for you.

Once a task is scheduled, be sure to click Run Now to trigger an immediate test run. This allows you to verify that everything is working as expected.

How to Know It Is Working

There are three things worth checking during the first couple of weeks:

Did It Actually Run?

This sounds obvious, but the first failure mode for a new scheduled task is often that it does not run when expected. Common causes include the wrong timezone, incorrect day-of-week assumptions, or cadence settings that did not save correctly.

Check that the output is there before you need it for several days in a row before you fully trust it.

Is the Output Still Good When Nothing Is Happening?

Pay particular attention during quieter weeks.

A task that produces a sharp, useful briefing during a CPI or Fed week but generates vague, filler-heavy output during slower periods may need better instructions. Just as a skill description sometimes needs refinement to trigger reliably, a scheduled task may need adjustments to maintain quality.

Do You Still Need It?

This sounds like a strange question to ask about something you just built, but scheduled tasks are easy to create and easy to forget about.

If you stop opening the file three weeks later, that is useful information. Perhaps the cadence is wrong, the format is not working, or the underlying skill needs to be revisited.

Retiring a scheduled task is just as valid an outcome as keeping one.

Where to Start, and What’s Next

If you already have the economic-calendar skill from the workshop, this is a 10-minute addition. Schedule it for weekday mornings, point the output to a file, and let it run for a week before judging the results.

If you do not have a skill built yet, start there first. Last week’s post walks through the process. Then come back to this once you have something worth automating.

This post intentionally left out two topics.

The first is connectors: what happens when the output of a scheduled task goes somewhere other than a file, such as Slack, email, or a shared document.

The second is live dashboards: artifacts that pull fresh data every time you open them rather than sending you a scheduled digest.

Both are natural next layers on top of “it runs itself,” and both will be covered in future posts.

For now, pick one thing you ask Claude for on a recurring basis and stop asking.

[Link to the post about skills for readers who have not built the economic-calendar skill yet]

That alone tells us something.

A year ago, many finance professionals were still skeptical about AI and viewed it as a nice-to-have. Today, they are showing up on their own time to learn how to use it because they recognize that AI proficiency is becoming a baseline expectation in the industry, not a differentiator.

While this expectation has not fully surfaced in financial services yet, it is already happening in technology. Many employees are being asked to upskill and incorporate AI into their daily workflows. I believe this trend will eventually extend to other industries, including finance.

The workshop focused primarily on Cowork and showcased many of the features that have made Claude so popular.

To me, Cowork represents the next level of AI proficiency beyond simple conversations. I’ve received feedback from many non-technical professionals that AI-assisted coding still feels intimidating. Cowork sits in the middle, helping non-technical professionals supercharge their workflows using AI without writing code.

In just 60 minutes, we covered a tremendous amount of material, including:

• Projects

• Connectors

• Skills

• Plug-ins

• Scheduled Tasks

• Live Artifacts

We also left out several popular features, including Dispatch, Claude in Excel, and Claude Design.

By the end of the workshop, participants had built a working economic calendar that automatically posts to Slack every weekday morning without lifting a finger.

If you’d like the full breakdown, this week’s paid post dives deep into how to build Claude skills from scratch, covering three different approaches, real examples, and what separates a skill that works reliably from one that breaks in production.

[Three Ways to Create Claude Skills (And What Makes Them Work)]

But first, let’s look at what’s happening in the broader AI and finance landscape.

The AI Provider Race for Financial Services

On May 5, Anthropic released an updated version of its Financial Services Plug-ins, expanding the offering to include ten specialized agents covering different aspects of work within the financial industry.

I’ve written two posts over the past few weeks about these plug-ins. If you’re interested, you can find those articles linked below to learn more about the plug-ins and how to customize them for your own workflows.

[A Practical Look at Anthropic’s Financial Services Plug-in]

[Making Claude’s Financial Services Plug-Ins Work for You]

Coincidentally, on the same day, Perplexity launched Computer for Professional Finance.

It is an AI research and analysis platform built around the workflows finance teams run every day. Teams with licensed data subscriptions can connect through MCP connectors to providers such as Morningstar, PitchBook, and Daloopa. Teams without terminal subscriptions can use Computer’s built-in finance tools, which draw from 14 financial data providers.

One particularly interesting feature is transparency. Every number links back to its source. Users can see where data came from, when it was retrieved, and, for SEC filing-derived calculations, the exact math and source filing page used.

The platform is available today through Microsoft Teams, with Excel integration coming soon.

OpenAI also entered the race. On May 15, OpenAI launched personal finance tools for ChatGPT Pro subscribers. Through an integration with Plaid, users can connect accounts from more than 12,000 financial institutions, including Schwab, Fidelity, Chase, Robinhood, American Express, and Capital One, and query their finances using natural language.

Competition among AI providers for financial services clients is accelerating rapidly.

For industry professionals, it is worth paying attention to where each provider is investing and how they are positioning themselves.

The adoption stories are becoming increasingly tangible.

Earlier this year, Consumer Bank signed a multi-year agreement with OpenAI. OpenAI engineers will work directly with the bank to automate lending and onboarding workflows. According to the bank’s leadership, commercial loan processing times could be reduced from 30-45 days to roughly seven days, while opening complex commercial accounts could fall from more than a day to approximately 20 minutes.

Another compelling example comes from Norway’s sovereign wealth fund, NBIM.

NBIM has been using Claude since November 2024 to screen every new equity investment within 24 hours of entry for indicators of forced labor, corruption, and fraud. Approximately half of the fund’s 700 employees now build their own Claude-powered tools internally. The fund reports roughly 20% time savings, and CEO Nicolai Tangen has emphasized that sustainability and governance are inseparable from financial performance.

The broader trend is clear.

Every major AI provider now views financial services as a strategic market.

The race is no longer about which model is smartest.

It is about which provider can embed itself most deeply into the operational workflows of financial institutions and become part of the industry’s infrastructure.

For practitioners, the more important question is simple:

Which of these tools actually fits your workflow, and are you building familiarity with them today?

The Goldman Sachs Playbook: Built for AI, Not Retrofitting It

Of everything happening in AI and finance right now, Goldman Sachs may be the most instructive case study.

According to an article published by Klover AI, Goldman has taken a fundamentally different approach from many organizations.

Most firms are retrofitting AI into existing systems. They add a chatbot here, an automation layer there, and attempt to fit AI into infrastructure that was never designed for it.

I’ve also observed organizations adopting AI in a fragmented way. One department purchases ChatGPT, another adopts Claude, and another experiments with a different platform entirely. While experimentation is healthy, the lack of a coordinated strategy often limits knowledge sharing and reduces organizational leverage.

Goldman took a different path.

The firm built a centralized AI platform, GS AI Platform, secured within its own infrastructure. The platform houses multiple models, including GPT, Gemini, Claude, Llama, and internal models, all operating within Goldman’s environment.

What impressed me most is the decision to support multiple models rather than committing to a single provider.

In investment management, firms often subscribe to multiple market data vendors to reduce operational risk. If one data feed fails during market hours, another is available. Goldman appears to be applying a similar philosophy to AI.

This flexibility may prove increasingly valuable as the AI landscape continues to evolve.

Speed is another underappreciated advantage of this architecture.

Because the platform is centralized and purpose-built, Goldman can introduce new models and capabilities in weeks rather than months or years. When a better model becomes available, it can be added quickly. When a new use case emerges, teams can build for it immediately.

The platform routes prompts to the model best suited for the task, maintains auditability, and allows the firm to switch providers without retraining employees.

The customization layer is equally important.

Goldman does not provide every employee with the same generic assistant.

Access is tailored by role, department, and use case.

Research analysts receive access to research archives and financial datasets. Developers receive access to codebases and development tools. Specialized assistants support specific workflows, including Banker Copilot for investment banking and dedicated research assistants for analysts.

The results are meaningful.

Goldman reports productivity improvements ranging from roughly 20% to as high as 55% for certain development tasks. The GS AI Assistant has generated gains of up to 25% in document analysis and content creation workflows. Banker Copilot is reportedly reducing M&A preparation time by as much as 40%.

The article also compares Goldman Sachs with other major banks, including JPMorgan, Citi, and Bank of America.

Each institution has chosen a different strategy.

JPMorgan emphasizes proprietary control. Morgan Stanley relies heavily on external partnerships. Goldman focuses on flexibility through a centralized, multi-model architecture.

None of these approaches are necessarily wrong.

However, Goldman’s model is particularly interesting because it is designed for a world where the best AI model may change every few months.

What This Means for Everyone Else

You might argue that Goldman Sachs is not the right benchmark.

Most of us do not have Goldman’s resources, engineering talent, or budget.

But the underlying principles are accessible at any scale.

Build for flexibility rather than lock-in.

Understand which tools perform best for different tasks.

Create workflows that work consistently without requiring you to repeat instructions every time.

These are not enterprise-only concepts.

They are the same principles behind building an effective Claude skill on a $20 monthly subscription.

The investment professionals who attended last week’s workshop already understand that AI proficiency is no longer a future skill.

It is a present-day skill.

The question now is how to build that proficiency intentionally rather than acquiring it piecemeal when it becomes unavoidable.

Today, we’ll start with Skills.

In this post, we’ll cover what a skill actually is, three ways to build one, and what separates a skill that works consistently from one that falls apart in practice. I’ll use the morning-note skill from Anthropic’s Financial Services plugins as the running example because it is short, concise, and has a very clear output.

What Is a Skill?

A skill is not a saved prompt.

A prompt is one-time. A skill is a saved workflow designed for repetitive use.

In a more technical sense, a skill is a folder that, at minimum, contains a SKILL.md file. A .md (Markdown) file is simply a plain text file that you can open in Notepad or any text editor. For skills, it is where instructions and workflows are written and stored in plain English.

You can also bundle scripts, reference documents, and asset templates with a skill, but the SKILL.md file is where all the important decisions live.

The reason skills work consistently is something Anthropic calls progressive disclosure.

Claude reads the metadata block first, before the conversation even starts. That metadata tells Claude when the skill should activate. The full instructions load only when Claude decides the skill is relevant. Any linked files load only if Claude needs them. This keeps token costs low and behavior predictable.

The practical result is simple: you teach Claude your workflow once, and it applies that workflow every time, without you having to re-explain your preferences, format, or process in every session.

Unlike some other Claude features, skills can be used across multiple Anthropic products, including Chat, Cowork, and Claude Code.

What Should Be Built as a Skill?

According to Anthropic’s The Complete Guide to Building Skills for Claude, there are three common categories:

1. Document & Asset Creation

For example, a DCF-model skill that creates an Excel-based discounted cash flow model for a financial security.

2. Workflow Automation

Step-by-step instructions for a repeatable workflow.

We’ll use the morning-note skill as an example later in this post. The morning-note skill defines the process for creating a daily morning note.

3. MCP Enhancement

Instructions that guide Claude to use specific MCP connectors to enhance a workflow.

For example, you can instruct Claude to use a FactSet MCP connector to retrieve financial data needed for the morning note

The Three Ways to Create a Skill

1. Browse and Install from the Marketplace

This is the fastest path.

In Cowork, go to:

You’ll see a library of pre-built skills that can be installed with a single click.

When to use this:

When someone has already built exactly what you need.

For example, the Financial Services plugin comes with production-grade skills already connected to FactSet, Bloomberg, Morningstar, and other data sources. No configuration is required.

The limitation is obvious: you only get what already exists.

If your workflow requires a specific report format, niche data source, or firm-specific convention, you’ll likely need to customize or build your own skill.

2. Create with Skill Creator

This is the method I demonstrated during the workshop and the best starting point for most users.

In Cowork, go to:

Skill Creator is itself a skill. It acts as a conversational guide that asks questions and generates the SKILL.md file for you.

Here is the prompt I used during the workshop:

Help me create a skill called “economic-calendar” using Skill Creator. The skill searches MarketWatch, Investing.com, and Briefing.com for the complete U.S. economic calendar for the current week. For each release: event name, date and time ET, prior, consensus, why it matters, and what number would surprise markets. Group by day, chronological order, written as a macro analyst briefing a portfolio manager. Ask whether to display as a live artifact or post to Slack.

Skill Creator walks you through:

• Naming the skill

• Writing the trigger description

• Defining data sources and connectors

• Specifying the output format

Anthropic’s guide estimates 15 to 30 minutes to create a first skill, which aligns with my experience from the workshop.

Keep in mind that you must save the skill before you can use it.

Click Save Skill and choose whether to save it locally or in Claude.

Once saved, you can invoke the skill by typing:

• “Run my economic calendar”

• /economic-calendar

in any Cowork session.

3. Write a Skill from Scratch

This approach is for advanced users who need full control.

A skill folder has a simple structure:

A few naming rules matter:

• Folder names must be in kebab-case (morning-note, not Morning Note)

• The file must be named exactly SKILL.md

• The filename is case-sensitive

In our example with the morning-note skill, the folder looks like the following:

The Anatomy of a SKILL.md File

The best way to understand a SKILL.md file is to look at a real example.

The morning-note skill from Anthropic’s Financial Services plugins helps investment analysts summarize overnight developments and prepare for morning meetings.

The Metadata Block (YAML Frontmatter)

This is the most important part of the file.

It tells Claude whether the skill should be loaded before the conversation even begins.

The metadata sits at the very top of the file:

Here is the morning-note example:

Notice what the description does.

It explains:

• What the skill produces

• Who it is for

• Which phrases should trigger it

According to Anthropic’s guide, the description should include both what the skill does and when it should be used. Leave out either one, and the skill may not behave as expected.

Workflow Steps

After the frontmatter comes the workflow itself.

For the morning-note skill, the workflow contains four sequential steps:

Step 1: Overnight Developments
Scan earnings, news, macroeconomic data, and market context.

Step 2: Morning Note Format
Follow a predefined template with sections such as:

• Top Call

• Overnight Developments

• Key Events Today

• Trade Ideas

Step 3: Quick Takes on Earnings
Generate a structured summary table when a company reports earnings.

Step 4: Output
Format for Slack or Word distribution, with a maximum length of one page.

Each step contains additional structure.

For example, Step 1 is broken into:

• Earnings and Guidance

• News and Events

• Market Context

The more specific the workflow steps, the more consistent the output becomes.

The Domain Knowledge Block

This is where the skill becomes truly useful.

The morning-note skill includes an Important Notes section containing guidance such as:

Be opinionated. Morning notes that simply summarize news without a view are not useful.

Lead with the most important development. Don’t bury the headline.

“No news” is a valid morning note. Say “nothing material overnight, maintaining positioning.”

If you’re wrong, acknowledge it in the next morning note. Credibility matters more than being right every time.

This is what separates a skill from a prompt.

These are practitioner rules.

Without them, Claude generates a neutral news summary.

With them, Claude produces something that sounds like it came from an analyst who has actually run a morning meeting.

What Makes a Good Skill vs. a Bad One?

The “bad” example is not necessarily bad.

It is often just the first draft.

Skill development is an iterative process. You’ll continually refine and customize the workflow based on your needs.

For example, you might:

• Restrict the skill to use a specific MCP connector

• Add macroeconomic analysis

• Include firm-specific reporting standards

• Introduce additional validation checks

There is almost always room for improvement.

How to Know a Skill Is Working

Before considering a skill production-ready, run three tests.

Test 1: Trigger Accuracy

Run 10 to 20 queries that should activate the skill.

Anthropic recommends a trigger rate of 90% or higher.

If activation rates are lower, add more specific trigger phrases.

Test 2: False Positives

Run queries that should not activate the skill.

This should not trigger the morning-note skill.

If it does, the skill description is too broad.

Test 3: Consistency

Run the same request across several sessions.

If the output structure remains consistent, the workflow is well defined.

If the output varies significantly, parts of the instructions are likely ambiguous.

How Do You Know a Skill Was Invoked?

For Chat and Cowork users, there are several ways:

• Ask Claude directly which skills were used.

• Ask Claude during the conversation whether a skill is active.

• Watch the conversation itself. Claude often references the skill it loaded.

• In Cowork, active skills appear in the context panel on the right-hand side.

This makes it easy to verify which skills are currently being used.

Where to Start

Skills are not prompts you save.

They are workflows you delegate.

In this post, we covered:

• What a skill is

• Three ways to create a skill

• How Skill Creator can help build an economic-calendar skill

• The structure of a SKILL.md file

• Why the morning-note skill is an effective example

I hope this gives you a practical starting point for building your own Claude skills and using Claude to make your daily workflows easier to manage.

It has created a lot of buzz in the investment world, and rightly so. But this is just the beginning. It is an example of how powerful Claude can become when paired with the right toolbox. The text-based files require no coding, which makes it easy for non-technical professionals to use and customize the plug-in for their own needs.

Today, we will walk through the structure of Anthropic’s plug-in system and show a quick example of a customized solution.

Getting Started: How to Install the Plug-In

Before we dive in, let us get the plug-in installed. The Financial Services plug-in is not currently part of Claude’s built-in marketplace, so we will need to add it manually.

Step 1: Note the URL of the Financial Services plug-in

Step 2: Open Claude Desktop and navigate to the Cowork tab

Step 3: Go to:

Step 4: Paste the URL into the box and click Add

Once this is done, you will see all available agents ready to use. To add a specific agent, simply click the “+” sign next to it.

How to Use the Agents

Think of each agent as a real person with a specific job title and a set of skills. A junior investment banking analyst, for example, knows how to build a DCF model and prepare a pitch deck. In the same way, each Claude agent comes loaded with skills designed for its role.

Once you install an agent, such as the Earnings Reviewer in our example, you will see it includes skills such as:

• /earnings-analysis

• /earnings-preview

Most of the time, Claude will automatically select the right agent and skill for your task. But if you want to be more precise, here are a few ways to take control.

To invoke a specific skill

Or:

To invoke a specific agent

Or:

Understanding the File Structure

Now that we know how to use the plug-ins, let us look at how they are built.

Plug-ins are made up of plain text files, mostly markdown files (.md) and JSON files (.json). It may look a little overwhelming at first, but there is no code involved. These are simple text files written in plain English.

A .md file (markdown) is a plain text file you can open in Notepad or any text editor. In our case, it is simply where instructions are written and stored in plain English.

A .json file is also a plain text file, but it follows a specific structure with key-value pairs inside curly braces. In our case, it stores the plug-in name and version.

Here is the folder structure for the Earnings Reviewer agent:

There are three key files to understand:

How to Customize the Plug-In

The easiest way to get started is to download the repo directly from GitHub. Click the “Download ZIP” button to save the zipped folder, unzip it, and copy the agent folder you want to modify. In our example, that is the earnings-reviewer folder inside agent-plugins.

Here are the two most practical customizations you can make.

1. Swap the Data Connectors

Inside earnings-reviewer.md, the tools: line specifies which data providers the agent uses to pull financial data.

If you do not subscribe to the default providers, you can replace them with MCP connectors you have already set up in your Claude account.

For example, I have Financial Modeling Prep and Financial Datasets connected. Here is the updated tools and workflow section, replacing the yellow highlights shown in the screenshot.

2. Modify or Add a Skill

To modify an existing skill, open the SKILL.md file inside the relevant skills folder and edit the instructions directly in any text editor.

To add a new skill, create a new folder inside skills/ with its own SKILL.md file. You can use Claude’s built-in Skill Creator to help build it.

Once the skill is ready, add its name to the skills list in earnings-reviewer.md so the agent knows to load it.

Upload Your Customized Plug-In

Once you have finished your edits, zip the modified folder and upload it to Claude:

That is it. You now have a customized Earnings Reviewer agent using your own data connectors and tailored skills.

The most interesting part of this framework is not just the Financial Services plug-in itself, but the broader idea behind it. We are moving toward a future where domain experts can build lightweight AI agents tailored to their workflows without writing traditional software.

For investment professionals, this opens up a practical path to creating internal research assistants, customized workflow automations, and specialized analytical tools that align with existing processes and data sources.

The barrier to entry is becoming much lower than many people realize.

Before we get into the details, let us talk about what a plug-in is. A plug-in is a set of instructions that tells Claude to work as a subject matter expert in a specific field. The best part? No code is needed. Everything lives in plain files. There is nothing to build and no infrastructure to set up. Even a non-technical person can open and modify those files to customize a plug-in for their own use.

Now let us take a look at the Financial Services plug-ins, which introduced ten agents in the latest release:

• Pitch Agent

• Meeting Prep Agent

• Market Researcher

• Earnings Reviewer

• Model Builder

• Valuation Reviewer

• GL Reconciler

• Month-End Closer

• Statement Auditor

• KYC Screener

Each agent comes equipped with a set of skills, and some agents include connectors for live financial data. Below is a summary of the skills and MCP connectors assigned to each agent.

One important note: while the connectors are already configured in the plug-in files, most of them are enterprise-level data providers that require an active subscription.

Here is the full list of financial data connectors currently available:

How do I install the plug-ins?

To install the Financial Services plug-ins in Cowork, navigate to:

Customize → Personal Plugins → “+” → Create Plugin → Add Marketplace

Paste the GitHub repository URL and click Sync.

How do I use the plug-ins?

To use a specific agent, simply mention it in your prompt. For example:

"Use the Pitch Agent. Here is my target company list."

"Run this through the KYC Screener."

How do I use the skills?

To invoke a skill, simply use the slash command “/” or click the “+” button in the chat window.

A lot of people are excited about these Financial Services plug-ins, and rightly so. But it is worth setting the right expectation: this is not a one-size-fits-all solution. What Anthropic has released is a framework, a well-designed starting point that shows how Claude can be integrated into financial workflows.

To really feel the power, the key is customization. For example, the investment-related agents above are pre-configured with FactSet, Daloopa, and S&P Capital IQ connectors. Not every finance professional subscribes to those services. To make an agent work for your setup, you would need to open the agent’s .md file and swap in the connectors you actually have access to. You can also add new skills to any agent to expand what it can do.

Anthropic provides the framework. The real work is in making it yours.

With the same concept in mind, Anthropic also introduced plug-ins for Legal and Small Business using the same file-based design. This feels like the beginning of a broader pattern across industries.

If the pace of new releases feels overwhelming, here is a reframe: plug-ins are just files. They are readable, editable, and entirely within your control. Instead of waiting for Anthropic to build the perfect plug-in for your workflow, consider building your own. You know your process better than anyone. That is exactly the kind of customization this framework was designed to support.

Explaining a workbook

Claude can scan a workbook, provide a summary of what it does, and explain formulas with complex cell references. This is extremely helpful when you have inherited or been handed a complex spreadsheet and need to modify or enhance it. There are still many complex financial models built using spreadsheets, and this feature can meaningfully increase productivity.

Debugging and fixing errors

With complex spreadsheets, you will often run into circular references and broken links. Claude being able to debug and fix errors is extremely helpful. I used to work with a real-time trading model built entirely on a spreadsheet, using many in-house add-ins and customized functions. It was extremely stressful to debug a circular reference or fix a broken link during trading hours. Having Claude handle this would have saved a lot of time and anxiety.

Importing financial data effortlessly

In the past, the quant team had to build various tools to pull financial data from a database, or we used the Bloomberg Excel tool or similar tools to import financial data. It was often slow or inefficient. I still remember running into usage limits when using Bloomberg Excel to import financial data, ending up with a spreadsheet filled with #N/As. With Claude’s help, we can now use MCP connectors to import data directly into Excel. This is extremely helpful.

Using Skills or plug-ins from the Claude library

This is extremely helpful, as Claude automatically filters the Skills or plug-ins that are compatible with Excel so we can easily use them when building a spreadsheet. We can even ask Claude to create new Skills directly from the chat window within the spreadsheet.

Multi-tab model building

I tested this by asking Claude to build a workbook for portfolio analytics, and it was able to execute the request and produce multiple linked tabs properly.

Working with cell references

I also asked Claude to make the workbook dynamic using cell references, offsets, and other functions. When pulling financial data, the number of data points returned can vary due to trading days and calendar conventions. Making the workbook dynamic produces a more versatile result that does not need to be updated frequently.

Handling complex calculations

I asked Claude to produce a workbook that runs simulations for a given stock. It calculated the parameters using historical data and created simulated time series across multiple tabs, making it very easy to follow the logic and review the results.

Producing outputs in different formats

Claude can extract information from an Excel workbook and create PDF or Word documents. This is very helpful if you want to create external-facing communications using the workbook as a reference.

There are also a few things Claude cannot do. It is not able to manipulate add-ins or create VBA macros directly. For example, if you want to use Solver or any other add-in, Claude will not be able to run it for you. Instead, it will provide detailed steps and all the relevant cell references so you can run it yourself. It also cannot install VBA macros, so it will not create buttons automatically. Instead, it produces the VBA code and instructs you to copy, paste, and assign the macro yourself. It is not perfect, but I find it very useful overall.

One more important note: always inspect the outputs carefully. Claude can make mistakes, so it is still better to double-check. One approach, similar to software engineering, is to produce a few test cases and ask Claude to validate the calculations against them.

Version control can also be tricky. Similar to AI-assisted coding, once changes are saved, they cannot be reversed unless version control is built in. It is very important to duplicate the workbook or ask Claude to create a snapshot before making any changes. Claude does have a feature that produces a log of all actions it performs, which is helpful for tracing changes.

Overall, Claude in Excel works quite well in my opinion. I have created a separate paid post for subscribers with detailed instructions, including the exact prompts, setup steps, and what I found along the way. For those interested, the detailed guide is available here: [Claude in Excel: Step-by-Step Setup for Every Feature]

Before You Start — Installation and Plan Requirements

Claude for Excel requires a Claude Pro subscription at minimum ($20 per month). It is also available on Max, Team, and Enterprise plans.

To install the add-in:

1. Open Excel

2. Go to Insert > Add-ins > Get Add-ins

3. Search for “Claude”

4. Click Add and sign in with your Claude account

5. The Claude sidebar will appear on the right side of your screen

Feature 1 — Model Selection (Sonnet vs Opus)

The model selector sits at the top of the Claude sidebar. Three options are available:

How to set it:

1. Open the Claude sidebar

2. Click the model selector at the top

3. Choose your model based on the task:

   • Sonnet 4.5 — fastest, best for targeted tasks like fixing a formula, answering a quick question, or making a small edit

   • Opus 4.6 — more powerful, better for complex multi-tab builds, connected data analysis, and heavy calculations

   • Opus 4.7 — the most capable model, best for the most demanding financial modeling tasks requiring the highest level of reasoning and accuracy

4. You can switch models mid-session if the task changes

The practical rule: use Sonnet for questions, Opus 4.6 for standard builds, and Opus 4.7 when the analysis is complex enough that output quality matters more than speed.

What to watch for: Opus 4.7 is the slowest of the three. For most portfolio analysis work, Opus 4.6 is the right balance. Reserve Opus 4.7 for your most demanding tasks such as multi-phase builds pulling live data across many tabs simultaneously.

Feature 2 — Global Instructions

The most important setup step in this entire post. Global Instructions let you define your working standards once and have Claude follow them automatically every session.

What it is: a persistent instructions field in the Claude sidebar that applies to every conversation in Excel.

How to set it up:

1. Open the Claude sidebar in Excel

2. Click the Settings icon at the bottom of the sidebar

3. Find the Instructions field

4. Paste your instructions and click Save

What to include for financial work? Feel free to use this prompt as your starting point:

“Always use IB formatting — blue text for hardcoded inputs, black text for formula-driven cells. Never modify, rename, or delete any existing tab unless I explicitly ask. Before making any significant changes to an existing tab, copy its current state into a new tab called ‘Snapshot_[tabname]_[date]’ and do not touch that snapshot tab under any circumstances. Default to Financial Modeling Prep for all financial data requests unless I specify otherwise. At the end of every response, tell me which Skills and connectors you used, if any. Always maintain a running log of every action you take in a tab called ‘Claude Log’. If it does not exist, create it at the start of every session.”

Feature 3 — Version Control and Session Logging

What it is: a combination of built-in logging and external habits that together give you the closest thing to version control available in Claude for Excel today.

The honest reality: there is no native rollback. Once Claude edits your workbook and you save, those changes are permanent.

Option 1 — Save to OneDrive

1. If you have access, move your workbook to OneDrive or SharePoint before your session starts. Excel automatically captures version history for files stored there

Option 2 — Enable Session Logging

1. Click the Settings icon at the bottom of the Claude sidebar

2. Find Session Logging and toggle it on

3. Claude automatically creates a Claude Log tab in your workbook

4. Every action Claude takes during the session is recorded there in order

5. The log tab stays in the workbook when you save — it persists after the session ends and is there when you reopen the file

Option 3 — Snapshot Before Major Changes

Ask Claude creates a snapshot tab before any significant edit without being asked use the following prompt as an example.

“Before making any changes, copy the current state of [tab name] into a new tab called ‘Snapshot_[tabname]_[today’s date]’.”

What to watch for: chat history does not persist between sessions. Every time you close Excel, the conversation resets. At the start of each new session, orient Claude by asking it to read the workbook and summarize what has been built before continuing.

Feature 4 — Explaining a Spreadsheet and Working with Cell References

What it is: Claude reads your entire workbook at the cell level — every formula, every dependency, every tab — and explains it in plain English with clickable cell citations.

How to use it:

1. Open the workbook in Excel with the Claude sidebar active

2. Make sure Claude can see the workbook. It reads whatever is currently open

3. Ask Claude to summarize the workbook, including what each tab does, what the key inputs are, and what the main outputs are

4. Ask follow-up questions about specific cells, formulas, or calculations

5. When Claude cites a cell reference in its response, click it. Excel navigates directly to that cell

How to make your workbook dynamic using cell references:

1. Ask Claude to identify any hardcoded values that should be driven by cell references instead

2. Ask Claude to replace them with formulas using OFFSET, INDEX, or dynamic range references

3. This is particularly useful when pulling financial data where the number of rows returned varies by date range or trading calendar

What to watch for: on very large workbooks with many tabs and complex nested formulas, Claude occasionally needs more than one pass to fully trace a dependency chain. If the first explanation seems incomplete, ask it to go one level deeper on the specific cell you are investigating.

Feature 5 — Error Debugging

What it is: Claude finds the source of broken formulas, circular references, and broken links, explains why they occurred, and proposes a targeted fix without touching the rest of the model.

How to use it:

1. When you see an error in a cell (#REF!, #VALUE!, #NAME?, or a circular reference warning), note the cell address

2. In the Claude sidebar, tell Claude the cell address and the error type

3. Ask Claude to find the source and explain what broke

4. Claude identifies the root cause and proposes a specific fix

5. Review the proposed change. Claude shows you exactly which cells will be affected before making any edit

6. Approve or reject the fix

Common error types Claude handles well:

• #REF! — from deleted rows, columns, or tabs that formulas still reference

• Circular references — common in financial models where interest expense feeds back into debt calculations

• Broken external links — from workbooks that have been moved or renamed

• #NAME? — from missing named ranges or functions from add-ins that are no longer active

What to watch for: on deeply nested formulas, Claude occasionally identifies a symptom rather than the root cause on the first pass. If the proposed fix does not resolve the error, ask Claude to look one level deeper at the cells driving the broken formula.

Feature 6 — Connectors (Live Financial Data)

What it is: Connectors let Claude pull live financial data directly into your workbook from external data providers — without leaving Excel or touching a browser tab.

How to enable a connector from within the Claude sidebar:

1. Open the Claude sidebar in Excel

2. Click the connectors icon in the sidebar

3. Browse the list of available connectors

4. Click on the connector you want to use

5. Follow the authentication steps if prompted

6. The connector is now active for your session

Example — Financial Modeling Prep: Financial Modeling Prep is one of the connectors available in the sidebar. Once enabled, Claude can pull live prices, earnings data, and company fundamentals directly into your workbook on demand.

How the permission system works:

• Claude notifies you every time it is about to call a connector

• A permission pop-up appears before any data is pulled

• You approve or reject each request

• Note Claude never pulls external data silently. Every call is visible and requires your confirmation

Note on connector availability: the connectors visible in your sidebar reflect what is available on your plan. Institutional providers such as FactSet, S&P Global, LSEG, and PitchBook require existing subscriptions and are available at the Enterprise plan level.

Feature 7 — Skills and Financial Plug-ins

What it is: Skills and plug-ins are pre-built workflows available in the Claude library that extend what Claude can do inside Excel. Claude automatically filters the library to show only Skills that are compatible with Excel. Non-Excel Skills are excluded from the list automatically.

How to access available Skills:

1. Open the Claude sidebar in Excel

2. Type / in the chat input

3. A menu appears showing all Skills and plug-ins compatible with Excel

4. Select one directly from the menu, or describe your task naturally and Claude applies the relevant Skill automatically

How to use a Skill:

• Select it from the “/” menu and describe your task, or

• Simply describe what you need — Claude recognizes when a Skill applies and uses it without being prompted

• To confirm which Skill Claude used, check its response — it will note the Skill applied if you have included that instruction in your Global Instructions

Creating new Skills from within Excel: You can also ask Claude to create a new Skill directly from the chat window inside the sidebar. Claude builds the Skill definition and saves it to your library, where it becomes available for future sessions.

Feature 8 — Multi-Tab Model Building

What it is: Claude builds and connects multiple tabs in a single session — creating each tab, populating it with data or formulas, applying consistent formatting, and linking tabs together with correct cross-tab formula references.

How to use it:

1. Tell Claude the full tab structure you want. Name each tab and describe what it should contain

2. Tell Claude the order to build in and how each tab should link to the others

3. Claude builds each tab in sequence, maintaining connections between them throughout

4. After each tab is built, review it before asking Claude to move to the next

What to verify after a multi-tab build:

1. Click into each new tab and confirm it populated correctly

2. Check cross-tab formula references. Click into a formula and verify the tab and cell references point to the correct locations

3. Open Formulas > Name Manager to confirm any named ranges Claude created are present

4. Check the Claude Log tab to review every action taken during the build

What to watch for: on very large builds spanning many tabs, Claude occasionally loses track of a cross-tab reference late in the session. If a formula in a later tab returns an error, ask Claude to trace the reference back to its source tab before making any changes.

Feature 9 — Complex Math Calculations

What it is: Claude handles advanced financial and quantitative calculations natively inside Excel, writing the formulas directly into your workbook without requiring add-ins, a Python environment, or external tools.

I tested two complex examples during my testing:

Portfolio risk metrics — I asked Claude to calculate weighted portfolio beta, annualized volatility, Sharpe ratio, and a full correlation matrix across multiple holdings. Claude built all of it natively using Excel’s built-in functions, formatted cleanly and linked correctly across tabs.

Monte Carlo simulation using Geometric Brownian Motion — I asked Claude to run 100 simulated price paths for a stock over 252 trading days. Claude calculated the historical parameters from raw price data, generated the full simulation across 100 columns, and produced a results tab with percentile outcomes, probability of gain, probability of loss, and a distribution chart — all in one session.

Both examples were built across multiple tabs so the inputs, calculations, and outputs were clearly separated and easy to follow.

Feature 10 — VBA Macro Writing and Solver Setup

VBA Macros

What it is: Claude writes VBA code that you install manually in Excel’s Visual Basic Editor. This is one of the features I mentioned in my free post as something Claude cannot do. Let me clarify that nuance here. Claude cannot install VBA macros automatically or run them for you. What it can do is write the full macro code, including error handling and comments, which you then install in one manual step. The result is a working macro that runs exactly as designed.

How to install a VBA macro Claude writes:

1. Copy the VBA code from the Claude sidebar

2. Press Alt+F11 (Windows) or Option+F11 (Mac) to open the Visual Basic Editor

3. In the left panel, find your workbook name and click on it

4. Go to Insert > Module

5. Paste the code into the module window

6. Close the Visual Basic Editor

7. Go to File > Options > Trust Center > Trust Center Settings > Macro Settings > Enable all macros

8. The macro is now installed and ready to run or assign to a button

Solver Setup

What it is: Claude cannot run Solver directly. Solver is a separate Excel add-in that operates independently of the Claude sidebar. What Claude does is build the complete optimization structure your model needs and provide the exact Solver settings to enter. You open Solver, confirm the settings, and click Solve.

How to enable Solver if it is not already active:

1. Go to File > Options > Add-ins

2. At the bottom, select Excel Add-ins from the Manage dropdown and click Go

3. Check the Solver Add-in box and click OK

4. Solver now appears under the Data tab in your ribbon

How the workflow works in practice:

1. Ask Claude to build the optimization structure, specify what you are optimizing, what the decision variables are, and what constraints apply

2. Claude creates the necessary cells, labels them clearly, and tells you the exact Solver settings to use

3. Open Solver via the Data tab

4. Confirm the objective cell, variable cells, and constraints match what Claude built

5. Select the correct solving method (GRG Nonlinear for most financial optimization problems)

6. Click Solve

Feature 11 — PDF and Word Export

What it is: Claude can extract information from your Excel workbook and generate formatted PDF or Word documents from it, without manual formatting or copy-pasting into another application.

How to use it:

1. Tell Claude which tab or data range to use as the source

2. Specify the output format — PDF or Word

3. Describe any formatting preferences — title, sections, what to include or exclude

4. Claude generates the document and makes it available to download

What works well:

• One-page summary reports and executive scorecards

• Structured tables pulled from workbook data

• Narrative summaries with embedded data references

What to watch for: complex Excel charts do not always translate cleanly into the exported document. Review the output before sharing anything externally.

Feature 12 — Saveable Custom Skills

What it is: once you find a workflow that works consistently, you can save it as a reusable Skill and run it with one click in any future session.

How to create a Skill directly from within Excel:

1. Type / in the Claude sidebar

2. Ask Claude to help you create a new Skill based on a workflow you describe

3. Claude builds the Skill definition and saves it to your library automatically

What to watch for: the more specific your Skill description, the more reliably Claude triggers it. Vague descriptions lead to inconsistent triggering. Include the specific task type, the input format, the output format, and any constraints — the same way you would write a good Standard Operating Procedure for a new team member.

Conclusion

Claude in Excel is a capable tool, but it rewards careful use. A few pieces of advice before you get started. Always verify the numbers. Claude can make mistakes, and a formula that looks right may not always be right. Build in a habit of checking outputs before they go anywhere near a client or a decision. Treat version control seriously. As covered in Feature 3, there is no rollback once changes are saved. Snapshot before you start, save to OneDrive, and enable session logging every time without exception. Finally, consider using test cases to validate complex calculations. Describe a scenario with a known answer and ask Claude to confirm the result matches. This is similar to how software engineers write unit tests, and it is one of the most effective ways to build confidence in Claude’s outputs before you rely on them.

Two Sigma on the Future of AI in Investment Management

Two Sigma published a two-part article series featuring senior executives discussing AI trends in investment management. The series includes perspectives from Matt Greenwood, Chief AI Innovation Officer; Ben Wellington, Head of Complex Feature Engines; Jin Choi, Head of Technique Forecasting; Mike Schuster, Head of AI Core; and Jeff Wecker, Chief Technology Officer.

There are several insights I would like to share.

For my fellow quant colleagues: in the future, instead of simply deploying AI models, AI will become the operating system for quantitative investing. Research will transition from “idea generation” to the rapid evaluation of ideas. I have been very interested in this topic, as I have discussed rapid prototyping and Paper2Code in past issues. LLMs are already enabling us to accelerate investment workflows that traditionally take a long time to process. As Matt Greenwood put it, “The future isn’t AI replacing humans; it’s humans who use AI well replacing humans who don’t.” Future quants will not be valued purely for idea generation and model building. They will also need to be expert practitioners of AI toolkits and agentic workflows to turbocharge their research process.

Another interesting trend: the era of “bigger is better” in AI seems to be winding down. The new emphasis is on efficiency. Scaling the right thing matters far more than pure scale. Beyond building the right model, interpretability, security, and alignment are becoming central priorities. The human-in-the-loop framework will not go away in the quant research space. Skilled quant researchers remain crucial, as rigorous research discipline and safe deployment require more oversight, not less.

It is encouraging to read these perspectives from practitioners who are actively navigating this shift inside one of the world’s most respected quantitative firms.

YC 2026 Finance Cohort: Four Startups Worth Watching

This week, beyond reading from institutional practitioners, I explored a few interesting AI startups in the financial industry from the Y Combinator portfolio.

Standard Signal is described as the first hedge fund that researches and executes trades purely with AI. As Standard Signal is a new addition to Y Combinator, there is not much public information available yet. While I remain skeptical, I am genuinely curious about the methodology behind a 100% AI-driven hedge fund. I am a strong believer in AI with a human in the loop, echoing Two Sigma’s position discussed above, so I will be watching this one closely.

Trata is an AI-powered research desk for hedge funds that uses agents to interview analysts and write investment research. While I have some reservations about the product, it raises an important point about conversations. As AI can increasingly manage and access everything digital, offline activities become more valuable. Conversations not only create human connections but also generate ideas and innovations. I am not entirely sure we should capture all of them using AI, but I do think it offers a competitive edge when deployed thoughtfully.

Finbar is described as the AI investment analyst used by several top-20 AUM hedge funds. It automates the collection and compilation of data that junior analysts traditionally gather manually. I am not a research analyst, so I cannot speak to the full depth of Finbar’s datasets, and there are other players in this space. That said, the feature that excites me most is its ability to consolidate all the data in Excel and allow analysts to adjust parameters and run scenarios.

This reminds me of Daloopa, which similarly collects and consolidates financial data and uses Excel as its primary product interface. For people outside finance, that may sound underwhelming. But I do not see Excel going away in finance anytime soon. It has been around for decades, and it is what senior management knows and trusts. It offers a simple interface with enough flexibility for them to manipulate data and get results without being overwhelmed. I still remember my first job on a market-making desk, where my manager ran live data feeds and a live interest rate curve builder entirely in Excel. He had hired a team of quants to build Excel-based analytics solutions precisely because he could manipulate the data and see results on his own terms.

Bayesline is the startup that interests me most. Founded by alumni from Bloomberg and BlackRock, it is a GPU-powered financial analytics suite for institutional investors. It appears capable of building custom factor models and thematic factors in real time, with full risk analytics support. The combination of institutional pedigree and GPU-native infrastructure is a differentiated approach worth watching. I would definitely like to learn more about it.

A Bigger Question: Can AI Startups Survive the Big AI Companies?

With the largest AI companies constantly rolling out new features, particularly those designed to serve the financial industry, I am growing concerned about the startup landscape. For some AI finance startups, their product offerings could easily be replicated by new capabilities from Claude, OpenAI, or Google. Startups may argue they have proprietary algorithms that differentiate them. But the point is this: major AI companies offer their models as platforms that make it relatively easy to replicate or reverse engineer those offerings.

And as a quant, how many times have I tested vendor solutions only to find they are not quite what I need? Now, tools like Claude and OpenAI allow me to build exactly what I want, at perhaps 20% of the time and cost it would have taken before. Should I still work with vendors, or simply build it myself?

I once spoke with the head of data science at a venture capital firm. She said exactly the same thing. For datasets based on publicly available online information, her team decided to build them in-house.

This brings me to the Exabel survey, which found that 94% of firms plan to increase their spending on alternative data this year. What I find more interesting than the headline number is where the budget is actually going. Traditionally, I would have expected most of that spending to go toward purchasing alternative data from vendors. But that is not what the data shows.

( source: https://hedgefundalpha.com/news/fund-managers-spend-more-on-ai-in-2026/)

A significant portion of the budget is being allocated to software, technology, and headcount, not just buying data. In fact, spending is relatively balanced across these categories, suggesting that firms are investing in the infrastructure and talent needed to build and manage data internally.

I believe the advance of AI is motivating firms to bring the creation and management of alternative data in-house, which supports the argument above. In that context, the startups that excite me most are those offering solutions that help companies customize the alternative data they want, rather than simply selling pre-packaged datasets. In the future, even data solutions will need to be bespoke.

There has been a lot of interesting AI news over the past few days. The most shocking one is the Claude Code leak, where the source code of Claude Code was accidentally included in one version of the download. According to a VentureBeat article, there are significant implications. The developer world has gone crazy downloading and analyzing proprietary features, such as an agentic memory architecture that solved “context entropy,” the tendency for AI agents to hallucinate as long-running sessions become more complex. The leak also gives competitors a chance to level the playing field by gaining access to its source code.

For those of us deploying AI in financial workflows or client-facing tools, the most important implication is the heightened security risk. Every actor who has access to the source code, including Anthropic’s competitors or malicious actors, now has a potential way to bypass security guardrails and permission prompts. To mitigate the risk: always download Claude Code using Anthropic’s native installer, avoid running Claude Code in unfamiliar environments, and rotate API keys while monitoring for usage anomalies.

A Shift to AI-Native Infrastructure

But while that story dominated developer circles, there was a quieter piece of infrastructure news that I think matters more for investment professionals.

At QCon London 2026, Morgan Stanley redesigned its entire API program for AI agents using MCP and an open-source framework called CALM, developed by FINOS. The headline number: the bank’s first API took two years to reach production. With the new architecture, it now takes just two weeks.

The demo also included a compliance guardrail, a denied-symbols list that blocked queries for specific financial instruments, where Claude was rejected in real time from accessing a restricted symbol. There was also a demo showcasing Google’s Agent-to-Agent protocol running alongside MCP: a portfolio rebalancer that discovered trading agents by skill and booked trades autonomously. That last detail is worth sitting with.

Personalization at Scale

I have been a strong advocate of using AI to democratize finance. As I mentioned in my previous blog post, I believe AI will reshape access to financial services in several meaningful ways. Smaller asset managers, startups, and independent financial advisors will be able to level the playing field and compete with large institutions. Everyday investors will have access to tools that genuinely improve their financial health. And instead of hiring multiple professionals across different specialties, financial advisor, accountant, insurance agent, we will eventually have a one-stop shop for all our financial needs.

I am happy to see real progress on this front. According to the Fortune article, Robinhood’s Strategies tool, trained on LLMs, no longer maps customers to one of twenty or so existing ETF baskets. It can now customize a portfolio for a customer who plans to buy a home. That is a meaningful shift.

The McKinsey report on Wealth Management in 2035 adds important context. It paints a picture of an industry under pressure from three directions at once: demographic change, AI capability, and a shifting definition of trust.

The next generation of investors are digital natives, fluent in social media and shaped by the attention economy. They are transforming what they expect from financial advisors, demanding transparency, immediacy, empathy, and shared values. At the same time, nearly 40 percent of financial advisors are expected to retire within the next decade, creating a significant capacity gap that technology-augmented models will need to fill.

According to the same report, more than 62 percent of independent advisors surveyed in 2024 used some form of AI for efficiency, but only about 20 percent used it for client-facing tasks. McKinsey expects this to evolve into agentic systems that reason, recommend, and act on behalf of advisors and clients, under human oversight.

With growing interest in private equity, real assets, commodities, and digital assets, investor portfolios will become broader and more global. Future investors will think about their overall household balance sheet, not just a portfolio. Future advisors will need to treat every client the way family offices treat their ultra-high-net-worth clients. Hyper-personalized financial advice will become the main trend, and AI is what makes it affordable at scale.

This is the version of democratization I find most exciting. Not AI replacing advisors, but AI equipping them to offer services that were previously reserved for the wealthy to a much wider range of people.

What stands out to me is how these stories connect. One highlights the risks that come with more powerful AI systems. The others show how infrastructure and applications are evolving to use them in practice. The real opportunity sits in between, where security, control, and innovation need to evolve together.

I was very honored to lead two ChatGPT workshops: AI Investment Edge with 100 Women in Finance and Work Smarter with AI with CFA Seattle. For those who attended, I really appreciated you showing up, and I am so glad the sessions were helpful.

The content was slightly different for each workshop, customized to be as relevant to the audience as possible. We covered many useful features in ChatGPT, such as Apps, GPTs, and Agent Mode. We talked about how to use these tools to streamline investment research workflows, create reports that compare financial data across multiple years, and even use AI to reverse engineer financial data.

With the overwhelmingly positive feedback from the two workshops, I am finally ready to launch my first online course: A ChatGPT Course for Investment Professionals.

The course covers many of the topics I taught in the workshops. I will slowly expand it with additional modules and continue updating the content with the latest changes.

You can use the link below to get early access. You will receive 90 days of free access while I continue building additional material.

👉 Get early access to A ChatGPT Course for Investment Professionals

Currently in the works:

• Claude for Investment Professionals

• A directory of MCP connectors for financial services

• Tips for using Claude and ChatGPT for Excel

• Practical prompts and workflows for investment professionals

👉 Get early access to the course

New Research from Anthropic: Labor Market Impact of AI

Observed vs Theoretical AI Usage

Anthropic recently published a research paper titled “Labor Market Impacts of AI: A New Measure and Early Evidence.”

In it, they introduced a new metric called observed exposure, which quantifies the percentage of real-world automated usage of tasks that LLMs could theoretically speed up in professional settings. They also introduced a theoretical coverage measure called the β score, which represents the full scope of what AI could do. The gap between the two is the central finding of the paper.

Implications for Finance

The research includes a figure comparing observed AI coverage with theoretical AI coverage across different industries. For business and finance, the gap is striking: theoretical coverage is high, but real-world observed usage remains a small fraction of what is possible. This tells us that AI is not yet being used anywhere near its full potential in our industry.

Source: https://www.anthropic.com/research/labor-market-impacts

Source: https://www.anthropic.com/research/labor-market-impacts

Anthropic also published a list of the most exposed professions, meaning those in which LLMs can theoretically speed up the largest share of tasks. Computer programmers rank first. For those of us in investment management, investment analysts show 57.2% theoretical exposure, meaning more than half of the tasks in that role could be made more efficient with AI. That is a meaningful number worth paying attention to.

OpenClaw and the Rise of Local AI Agents

What is OpenClaw?

Besides the exciting news in my own journey, there is also a lot of interesting news in the AI space surrounding OpenClaw.

OpenClaw is an open-source AI assistant agent that lives on your computer. It is not an AI model itself. To use OpenClaw, you subscribe to your own AI model, whether that is ChatGPT, Claude, Gemini, or another provider. What makes it fascinating is that it lives locally on your computer and can perform real tasks, such as running scheduled tasks, managing your emails, and even controlling your smart home system. The way you communicate with OpenClaw also feels more natural. It works through messaging apps like WhatsApp, Telegram, and Discord. You can text OpenClaw as if you are texting a friend, and it remembers the conversation.

Adoption and Popularity

Another interesting story is how OpenClaw has gained so much popularity in China. In line with China’s push for open-source AI, OpenClaw has been heavily marketed and promoted. There are stories about Tencent and other tech giants hosting events to help people install OpenClaw on their computers. There are, of course, many security risks behind this OpenClaw boom in China, but that is not the topic for today.

Real-World Examples

Here are some positive and negative examples of OpenClaw in action. It is always good to learn from both ends so we can make better decisions about incorporating it into our daily lives.

One story describes how OpenClaw went the extra mile to get a task done. The author wanted to reserve a table at a popular restaurant and tasked OpenClaw with it. It first tried to book online, but the restaurant was fully booked. Determined to complete its mission, it downloaded various software and voice apps so it could make a phone call and speak to a real person for the booking. In the end, the mission was accomplished. OpenClaw managed to talk to a human and found a table thanks to a last-minute cancellation.

On the other end, there is the story of Summer Yue, Director of Alignment at Meta Superintelligence Labs. She tasked OpenClaw with checking her inbox to see if anything could be deleted or archived, with explicit instructions not to take any action without her approval. However, the size of her inbox triggered a compaction event, causing OpenClaw to lose her safety instructions. It proceeded to delete her emails. She tried to stop it remotely from her phone, but the agent ignored her commands, and she had to physically run to her computer to shut it down.

Industry Response

With OpenClaw’s growing popularity, we have seen some notable responses from the broader industry.

Nvidia launched NemoClaw, an open-source stack that adds privacy and security controls to OpenClaw. I have to admit that I am a more conservative AI user, possibly because of my background in finance. I am not yet ready to give an AI agent full access to my computer. With the introduction of NemoClaw, I might be more willing to give it a try.

Another development comes from Anthropic. Anthropic launched Dispatch, a new feature within Claude Cowork that lets you control your desktop AI agent remotely from your phone. You can assign tasks to Claude from your phone and come back to completed work on your desktop.

Final Thoughts

Finally, on a personal note, beyond all the news this week, I am currently testing a few MCPs with various agentic features. I will share more once the verdicts are ready.

ChatGPT offers a powerful and accessible platform for doing exactly that. In this course, we explore how investment professionals can use ChatGPT to connect data sources, build reusable tools, and streamline research and analysis.

In this course, we will cover the following modules:

Module 1 – Understanding the ChatGPT Platform

Module 2 – Connecting Data with Apps and MCP

Module 3 – Building Custom GPTs for Reusable Workflows

Module 4 – Extending GPTs with Actions and APIs

Module 5 – Exploring Automation with Agent Mode

By the end of the course, you should have a clear framework for using ChatGPT as a practical tool to support research, automate routine tasks, and improve the efficiency of your investment workflow.

Module 1 – Understanding the ChatGPT Platform

Before we get started, let’s walk through some basic features of the ChatGPT platform.

Many people think of ChatGPT as a single tool, but in reality it is a platform composed of several components that work together.

Understanding how these pieces fit together will help you use ChatGPT much more effectively.

In the next few modules, we will explore several of these components in more detail, including:

• Apps

• GPTs

• Agent Mode

ChatGPT Platform Architecture

ChatGPT Platform

Core Chat Experience
│
├── Apps
│   Connect external tools and data
│
├── GPTs
│   Create reusable tools
│
└── Agent Mode
    Explore complex workflows

The diagram above shows the main building blocks of the ChatGPT platform.

Each component plays a different role in the system:

• Apps connect ChatGPT to external tools and data sources

• GPTs allow you to create reusable automation tools

• Agent Mode allows ChatGPT to perform complex tasks autonomously

Let’s walk through each of these components.

Apps

In the architecture diagram, Apps represent the integration layer.

Apps allow ChatGPT to connect to external tools, services, and data providers. Apps were previously known as connectors. They use a standard called MCP (Model Context Protocol).You may have heard about MCP in the news recently. Originally developed by Anthropic, MCP is a standard that allows AI systems to connect to external tools, data sources and services. This makes it easier for AI models to interact with software systems.

Why Apps are Useful:

Simple Example

One example is the Target App.

Using natural language, you could ask ChatGPT to search for products available at Target.

ChatGPT will display the results and direct you to the Target website if you decide to complete the purchase.

Work Example

Apps are also becoming available from financial data providers.

Several well-established financial data service providers now offer Apps, including:

• S&P Global

• MSCI

• Morningstar

If you subscribe to these services, you may be able to request access to their Apps. These integrations allow ChatGPT to retrieve financial data directly from professional data platforms.

In Module 2, we will demonstrate how to enable Apps from ChatGPT’s Apps store. We will also demonstrate how to connect to data providers that are not listed in the App Store but offer remote MCP servers for connection. This allows you to integrate additional data sources into your workflows.

GPTs

Next in the architecture diagram is GPTs, which represent the automation layer of the platform.

You will see this feature when you expand the left side panel in ChatGPT. GPTs allow you to create custom AI tools designed for a specific task.

In Module 3, we will walk through:

• How to use GPTs created by others

• How to build your own GPTs

Why GPTs Are Useful

GPTs work best for repetitive tasks, so you can think of them as utility tools.

Simple Example

For example, you could create a GPT that generates recipes for your family.

If family members have dietary restrictions, you can include those constraints in the GPT instructions.

Each week before grocery shopping, you could run the GPT to generate a meal plan.

Work Example

In a professional setting, you could create a GPT that:

• Performs web searches

• Gathers the latest news about specific financial securities

• Summarizes relevant developments

You could also create a GPT that:

• Runs portfolio analysis

• Generates dashboards and visualizations

• Prepares materials before client meetings

Later in Module 3, we will demonstrate how to build a GPT from scratch.

In Module 4, we will also show how to generate Actions, which allow the GPT to connect to external tools.

Agent Mode

Finally, we have Agent Mode, which appears in the architecture diagram as the automation and execution layer.

Agent Mode allows ChatGPT to perform multi-step tasks autonomously.

When enabled, ChatGPT can interact with websites and tools to perform tasks such as:

• Booking appointments

• Gathering information

• Creating presentations

It is capable of handling complex tasks from start to finish.

How I Use Agent Mode

I personally use Agent Mode as a learning tool.

If I am exploring a new idea and do not know where to start, I enable Agent Mode and ask ChatGPT to conduct research.

If the final result is useful, I then ask ChatGPT to:

• Document the workflow

• Translate the process into structured instructions

Those instructions can then be saved for later use.

In many cases, the workflow discovered through Agent Mode can eventually be turned into a custom GPT designed for repetitive tasks.

Module 2 – Connecting Data with Apps and MCP

One of the most powerful developments in the ChatGPT ecosystem is Apps.

Apps, formerly called connectors, allow ChatGPT to connect directly to external tools and data providers through a protocol called MCP (Model Context Protocol). This turns ChatGPT from a simple chatbot into a workflow hub that can interact with external systems.

For investment professionals, this is particularly useful because it allows ChatGPT to retrieve market data, financial information, and analytics directly from third-party providers.

What Apps Allow You to Do

Apps allow ChatGPT to connect to external services and retrieve information in real time.

Examples include:

• Market data providers

• Financial analytics platforms

• Document storage systems

• Research databases

• Brokerages and trading platforms

For example:

You can connect the Target app and use natural language to shop on Target’s website.

Instead of navigating the website manually, you could simply type:

“Find the best LEGO toy for a 4-year-old under $20 available on Target.”

ChatGPT will then query the Target system through the app.

While this example is retail-focused, the same concept applies to financial data providers.

Financial Apps Available in ChatGPT

Many financial data vendors now provide Apps that integrate directly with ChatGPT.

Examples include:

• S&P Global

• Morningstar

• MSCI

• LSEG

If you already have a paid subscription to these providers, you can connect their apps to ChatGPT to enhance your research workflow.

Once connected, ChatGPT can access the features exposed by the app provider.

For example:

The MSCI App allows users to retrieve information about MSCI indices, including index composition and methodology.

Instead of searching the MSCI website manually, you could ask:

“Show the current sector weights of the MSCI World Index.”

Affordable Data Apps for Smaller Firms

For small asset managers without access to expensive terminals like Bloomberg, several free or affordable data apps are available.

Below are some useful options.

Alpaca

Alpaca is a brokerage platform that focuses on algorithmic trading.

Its ChatGPT app allows users to retrieve:

• Ask Market Questions

• Access Live Trading Data

• Trigger Trading Actions

This can be useful for quick research or prototype analysis.

Example prompt:

“Retrieve the last 5 years of daily prices for AAPL and show the annualized return and volatility.”

TickerSage

TickerSage is a real-time technical charting tool for both crypto and supported equity markets.

Key features include:

• Real-time chart updates

• Advanced Technical Studies

• Market visualization tools

One of the most impressive features is that charts update in real time as new data becomes available.

Example prompt:

“Show me the real time chart for Bitcoin with RSI and EMA.”

Kraken

Kraken provides an app that focuses on crypto market intelligence.

It can generate:

• Market snapshots

• Top movers

• Trending crypto assets

• Market performance summaries

Example prompt:

“Generate today’s crypto market snapshot and show the top 3 movers.”

This is a quick way to monitor crypto market activity without opening multiple dashboards.

Connecting to External MCP Servers

Not all apps are listed directly in the ChatGPT App Store.

Some vendors provide remote MCP servers, which allow ChatGPT to connect directly to their systems.

Examples include:

• financialdatasets.ai

• Financial Modeling Prep

• Alpha Vantage

These providers offer APIs and datasets that can be accessed through MCP connections.

This allows ChatGPT to retrieve structured financial data such as:

• Financial statements

• Market data

• Company fundamentals

• Economic datasets

How to Connect Apps in ChatGPT

There are two main ways to connect apps.

1. Install Apps from the ChatGPT App Store

The easiest way to connect apps is directly from the left sidebar inside ChatGPT.

On the left panel, you will see an “Apps” section. This is where all available integrations live.

To install a new app:

1. Open the left sidebar

2. Click Apps

3. Select “Add more” to open the App Store

4. Browse available integrations and connect the ones you want

Once installed, the app becomes available to ChatGPT automatically. When a prompt requires data from that provider, ChatGPT can call the app in the background to retrieve the information.

2. Connect Remote MCP Servers (Advanced)

If a provider is not listed in the App Store, you can connect manually.

Steps:

1. Go to Settings

2. Enable Developer Mode in Advanced Settings

3. Return to the Apps section

4. Create a new app connection

5. Enter the MCP server URL provided by the vendor

This allows ChatGPT to communicate directly with the provider’s server.

How to Use Connected Apps

Once an app is installed, there are two main ways to use it inside ChatGPT.

1. Use the “+” Menu

You can access connected apps directly from the “+” button next to the message box.

Click the + icon, and you will see a list of available apps you have installed. From there, you can select the app you want to use for your query.

This method is helpful when you want to explicitly choose which app ChatGPT should use.

2. Mention the App Using @

Another way to use apps is to call them directly in your prompt.

You can do this by typing @ followed by the app name.

For example:

“@Alpaca show the latest quote and 1-year price chart for NVDA.”

This tells ChatGPT to route the request through the specified app.

In most cases, however, you do not need to manually select an app. If a relevant integration is available, ChatGPT may automatically use it in the background to retrieve the information needed to answer your question.

Why Apps Matter for Investment Workflows

For small investment teams, Apps dramatically expand what ChatGPT can do.

Instead of relying only on the model’s built-in knowledge, you can connect ChatGPT to live financial data and analytics systems.

This allows you to:

• Retrieve real-time market data

• Analyze financial statements

• Generate technical charts

• Monitor market activity

• Automate research tasks

In other words, Apps allow ChatGPT to become a lightweight research platform for asset managers.

Module 3 – Building Custom GPTs for Reusable Workflows

GPTs are custom versions of ChatGPT that you can create for specific purposes.

They help you build domain-specific specialists that assist with repetitive tasks.

For example:

• As a financial advisor, you can build a GPT to help you analyze and understand the portfolios of your clients.

• As a research analyst, you can build a GPT to analyze multiple years of financial statements and produce side-by-side comparisons.

GPTs can be helpful utility tools that streamline your workflow and improve productivity.

How to Build a GPT

To build a GPT, first navigate to the left side panel of ChatGPT and click on “Explore GPTs.”

Then you will find the “Create” button on the upper right corner.

This is where you can start building a GPT.

The GPT Builder Interface

After clicking the “Create” button, you will be taken to the GPT Builder.

On the left panel, you will see two tabs:

• Create

• Configure

The right panel is where you experiment with your customized GPT.

This allows you to test your GPT while you build it.

GPT Configuration

If you take a look at the “Configure” tab, there are several inputs you need to complete in order to make the GPT work.

Here are the key inputs.

Name:

The name of your GPT.

Example:

SEC Filing Comparison Assistant

Description:

A short explanation of what the GPT does.

Example:

A GPT designed to analyze and compare SEC filings across multiple years.

Instructions

This is where you provide specific instructions for the GPT to function.

To create the instructions, you can use the “Create” tab, where you describe:

• what you would like the GPT to do

• your expectations

• how the output should look

You can also include guardrails or rules to ensure the GPT stays aligned with your instructions and reduce hallucination.

Conversation Starters

Conversation starters are sample prompts that show the tasks this GPT can achieve.

These help users quickly understand how to interact with the GPT.

Example conversation starters might include:

• Compare Apple’s last three 10-K filings

• Summarize the latest quarterly report for Nvidia

• Analyze revenue trends across Tesla’s recent filings

Knowledge

The Knowledge section allows you to include additional context for the GPT.

You can upload up to 20 files.

When a file is uploaded, the GPT:

• breaks the text into chunks

• creates embeddings (a mathematical representation of the text)

• stores them for later retrieval

This process is called Retrieval Augmented Generation (RAG).

When you interact with the GPT, it can retrieve relevant information from the uploaded files to provide better responses.

Please note that images inside uploaded files are currently not processed.

Capabilities

Here are the capabilities currently offered.

Web Search : Allows the GPT to retrieve information from the internet.

Canvas: Canvas is a side-by-side interface that allows users to collaborate, edit, and refine writing or coding projects directly with GPT.

Image Generation: Allows the GPT to generate images.

Code Interpreter and Data Analysis: Allows GPT to analyze datasets, run calculations, generate charts, perform financial analysis

Actions

Actions allow the GPT to connect to external APIs and perform additional tasks.

This makes GPT significantly more powerful.

We will discuss Actions in more detail in Module 4.

Demo: Building a GPT to Compare SEC Filings

Let’s try to build a GPT that compares SEC filings across multiple years.

Here are the steps.

Step 1: Use the Create Tab

Use the “Create” tab to generate:

• the name

• the description

Step 2: Define the GPT Instruction

For this demo, the GPT should generate side-by-side comparisons for each item in 10-Qs and 10-Ks. The user should provide a ticker symbol as the input.

Note that it will automatically populate the input and fill in the details within the “Configure” tab. This process is typically iterative and may require a few rounds of refinement.

You should continue refining the instructions until you feel comfortable with the behavior of the GPT.

Step 3: Generate Conversation Starters

Use the Create tab to generate three prompt starters.

Example starters:

• Compare Apple’s last three 10-K filings

• Analyze changes in Nvidia’s recent quarterly filings

• Summarize revenue trends in Microsoft’s recent filings

Step 4: Enable Web Search

Lastly, enable the Web Search capability.

By doing so, the GPT will search online to find the required filings.

Example Output

Once the GPT is configured, you can test it by entering a ticker.

The GPT should generate a structured comparison across filings.

A Note on Data Sources

You may feel uncomfortable asking the GPT to search online and simply trusting that it will find the correct source.

Don’t worry.

In the next module, we will discuss Actions, which allow you to connect GPT directly to external APIs.

This gives you full control over where the GPT retrieves its data.

For example, you could connect GPT directly to:

• market data providers

• financial databases

• internal research systems

We will cover this in Module 4.

Module 4 – Extending GPTs with Actions and APIs

Actions allow you to connect your Custom GPT to external APIs, giving the GPT additional capabilities to perform tasks using real-world data.

For example, you can connect a GPT to:

• Google Drive

• Gmail

• Financial data providers

• Internal systems

If a GPT has access to Google Drive, it can retrieve files and use them as additional context when answering questions.

If connected to Gmail, you could create a GPT that acts as an email assistant, summarizing incoming messages and suggesting next steps.

In the financial industry, many data providers offer APIs for software development. These APIs allow you to retrieve structured financial data such as:

• company fundamentals

• financial statements

• SEC filings

• market data

• news

By connecting these APIs through Actions, your GPT can call specific functions defined in the API documentation and retrieve the data automatically.

Demo: Retrieving Financial Data via API

Let’s walk through a quick example.

Suppose we want our GPT to compare SEC filings between companies.

Instead of:

• uploading filings manually

• relying on web search

we can configure the GPT to retrieve the filings directly from a financial data provider.

This approach ensures the GPT retrieves information from reliable structured sources, rather than relying on unstructured web pages.

In this example, we connect the GPT to FinancialDatasets.ai.

The GPT can then retrieve financial data directly from the API and use it to perform analysis or comparisons. Below is an example of retrieving SEC filings through the API.

Notice the cURL command in the upper-right corner. This command shows the structure of the API request, which we can use to generate the schema for our Action.

url --request GET \ --url 'https://api.financialdatasets.ai/filings?limit=10' \ --header 'X-API-KEY: <api-key>'

Alternative APIs You Can Experiment With

FinancialDatasets.ai is just one example. There are several other free or affordable APIs that work well when testing Actions.

Some useful options include:

• Financial Modeling Prep

• Finnhub

• NewsAPI

These services allow your GPT to retrieve data such as:

• stock prices

• company fundamentals

• earnings history

• financial statements

• market news

Experimenting with these APIs is a good way to extend your GPT beyond static knowledge.

Creating Actions

To implement Actions, you need to provide a schema that defines how the GPT should interact with the API.

At first glance, the schema may look technical. However, you do not need to write it manually.

You can use ActionsGPT to help generate the schema automatically from the API documentation.

Once the schema is generated, the setup process is straightforward:

1. Paste the schema into the GPT Actions configuration

2. Add your API key for authentication

In general, Actions work best with APIs that support custom header authentication.

Example flow:

Paste API documentation
↓
Generate schema
↓
Copy schema

Here is the prompt that helps us generate the schema:

Create a spec for this API endpoint with custom header: curl --request GET \ 
  --url 'https://api.financialdatasets.ai/filings?limit=10' \
  --header 'X-API-KEY: <api-key>'

Once the schema is generated, copy and paste it into the Schema section of your Action configuration. In this example, we created a “GetFilings” action to retrieve SEC filings.

Lastly, we will need to provide the API key for authentication. You will click on “Authetication” then seleclt “API Key” and “Custom”.

You will input your API key and name your custom header “X-API-KEY”

Once this is done, you will be able to retrieve SEC filings.

Key Takeaway

Actions transform a GPT from a static assistant into a data-powered tool.

Once connected to APIs, your GPT can:

• retrieve live financial data

• compare company filings

• pull historical datasets

• automate research workflows

• integrate with external systems

For investment professionals, this means you can build custom research tools without writing traditional software.

Module 5 – Exploring Automation with Agent Mode

Agent Mode is one of the most powerful capabilities available in ChatGPT. However, it is currently not available on the Free tier or Go tier.

When Agent Mode is enabled, ChatGPT can interact with websites on your behalf.

For example, if you ask ChatGPT to find affordable hotels for an upcoming trip, you can watch, in real time, as it visits travel websites, clicks on links, and navigates between webpages to gather information and complete the task.

Unlike a normal chat response, Agent Mode performs the task step by step.

You can actually observe the model’s reasoning capability in action. It may ask itself questions, evaluate different options, and move back and forth between decisions until it finds a satisfactory result.

Because Agent Mode behaves similarly to a human performing the task manually, the process may take some time. It often starts from scratch, searching the web, opening websites, navigating pages, and gathering information along the way.

How I Use Agent Mode

Personally, I still do not feel completely comfortable letting AI make decisions for me, especially if it might perform actions that are not easily reversible.

Instead, I use Agent Mode primarily as a learning and discovery tool.

For example, I recently used Agent Mode to help plan part of my travel itinerary. I asked it to search for kid-friendly activities between San Francisco and Sonora.

I also added a few constraints:

• The locations must be close to the highway route we plan to take

• The stop must require less than a 10-minute detour

ChatGPT was able to successfully perform the task by searching for locations, evaluating options along the route, and identifying places that matched the criteria.

The prompt used for the travel planning task.

Example prompt:

Find kid-friendly activities between San Francisco and Sonora.

Requirements:
- Close to the highway route
- Less than a 10 minute detour
- Suitable for children

Finance Example: Comparing Analyst Recommendations

Agent Mode can also be useful for investment research.

A common task for analysts is to review and compare analyst recommendations for a stock. Normally, this requires searching multiple websites and reading several articles.

With Agent Mode, ChatGPT can browse financial news sources and compile the information for you.

Example prompt:

Research Nvidia (NVDA) and summarize the latest analyst recommendations.

Tasks:
- Find at least 5 recent analyst opinions
- Identify price targets and ratings
- Compare bullish vs bearish arguments
- Provide a short summary of overall market sentiment

Once the task begins, Agent Mode starts searching the web and visiting financial news websites. It may open articles from sources such as Reuters, CNBC, MarketWatch, or other financial publications.

It reads through the articles, extracts the relevant information, and compares the analyst views.

As it gathers information, the model evaluates different analyst perspectives. Some analysts may raise their price targets, while others may highlight risks or valuation concerns.

Agent Mode then organizes the findings and compares the viewpoints across sources.

Example Output

After reviewing multiple sources, ChatGPT produces a summary of analyst recommendations.

Turning Agent Workflows into Reusable Systems

For me, the most valuable step comes after the task is completed.

The next step is to ask ChatGPT to document its workflow.

For example:

Explain the workflow you used to find and compare analyst recommendations.

Translate this workflow into step-by-step instructions that could be used to build a Custom GPT.

This allows you to capture the process and convert it into a repeatable research workflow.

Instead of manually gathering analyst opinions every time, you can later build a Custom GPT designed specifically for monitoring analyst recommendations.

Key Takeaway

Agent Mode is not just a tool for automation.

It is also a powerful way to discover workflows that can later be turned into reusable prompts, templates, or Custom GPTs.

What Are Connectors?

Similar to apps in ChatGPT, connectors enable Claude to utilize external tools to complete tasks. Connectors normally utilize MCP (Model Context Protocol), a standard introduced by Anthropic that helps AI models connect to external tools, data, and services.

Where Can We Find Connectors in Claude?

To see the list of connectors currently available in Claude, you will need to go to the connector directory. If you use the web version of Claude, there are two ways to access the directory.

First, Claude recently introduced a new feature called “Customize.” You can find the “Customize” feature on the left sidebar. You will be able to see the connectors that you are currently connected to from there. If you click on the search icon, it will take you to the connectors directory where you can browse existing connectors available in Claude.

The second way to access the connectors directory is through Settings -> Connectors. From there, you can click on the “Browse Connectors” button in the upper right corner.

Here is a screenshot of the connectors directory:

How to Enable Connectors

It’s fairly simple. Once you find a connector that you would like to enable, simply click on the “Connect” button. If additional authentication is needed, Claude will take you to the authentication page for verification.

How to Use Connectors in Claude

It is fairly straightforward. Once you have a connector set up, Claude will automatically choose the connectors to use in order to complete the assigned task. To see what connectors are available for the chat window, simply click on the “+” sign and you can enable or disable connectors from there.

There is also a feature called “Tool Access,” where you can tell Claude how you want the connectors to be loaded. This is helpful when you have many connectors enabled.

If you have a particular connector in mind and would like Claude to prioritize using it, you can simply use “@” so Claude knows to prioritize that connector for a specific task.

Financial Data Providers Available as Claude Connectors

Here are some financial service providers that offer connectors in Claude. For most of the larger companies, you will need to be a subscriber to use this feature. I have included snapshots of the landing pages so you can see the available tools for each data provider. You can also find links to their documentation at the landing page to understand the full offerings.

S&P Global

FactSet AI-Ready Data

Morningstar

MSCI

Using Remote MCP Servers with Financial Data Providers

Besides the ones listed in the connectors directory, there are also many other data service providers that offer remote MCP servers. While this is not specific to Claude, remote MCP servers make it very easy to enable external tools with different AI models. Once you find the remote MCP server URL, you can easily integrate it into AI models.

Here is a quick snapshot of the Custom Connector in Claude. All you need is the URL to set up the connection. Here are three financial data service providers that I am able to connect to. I have also included the links to their remote MCP servers in case you would like to try them.

FinancialDataset.ai
Financial Modeling Prep
Alpha Vantage

Here is a snapshot of my financial data connectors working in full force. See how I asked Claude to retrieve the latest quote for AAPL. To complete the task, Claude automatically explores available tools for stock quotes. In this case, it picked Financial Modeling Prep and asked me for permission to proceed with data retrieval. This is a very powerful tool for investment professionals, as data accuracy is extremely important for investment-related tasks. Instead of having the AI model search online, we can ensure it retrieves data from credible sources.

Final Thoughts

I hope this post is helpful for investment professionals who are interested in using Claude in their workflows. I have been very impressed by Claude. Besides connectors, the skills feature is tremendously helpful, and the Claude Excel plug-in can not only help people understand legacy spreadsheets but also build spreadsheets from scratch.

Recently, I shifted my focus and decided to experiment more with using existing AI tools as someone without a technical background. This has been a fun journey. I’ve been playing with both ChatGPT and Claude. While this is still an ongoing research process, I want to share some interesting utility tools in ChatGPT that can be helpful for investment professionals. I will focus on ChatGPT first and share about Claude in the next blog post.

ChatGPT App Store

I previously shared about building customized GPTs and creating actions to connect to external APIs to extract financial data. With more apps available, there are now more options for retrieving financial data. There are apps from established financial data providers such as S&P Global and Morningstar.

Where to Find the App Store

To find the App Store, go to the sidebar and click on the Apps icon. You will be directed to the App Store.

From there, you can scroll to find the app you would like to connect or use the search feature in the upper right corner.

Once you find the app you want to connect, click the “Connect” button. For apps that require authentication, you will be directed to a page to provide your login credentials.

How to Use the Apps

There are two ways to use connected apps:

1. Click on the + button and select the app available to you.

   

2. Mention the app using @ in your prompt.

Examples of Financial Apps

Here are a few existing apps currently with no access restrictions.

Alpaca
Alpaca is a brokerage firm that specializes in algorithmic trading. It has an app in ChatGPT that allows you to retrieve historical stock prices, show the latest quotes, and provide market snapshots.

Example prompt:

TickerSage
TickerSage appears to be relatively new. It is a real-time technical charting tool for crypto and supported stock markets, with advanced technical indicators and charting features. The most impressive part is that the chart updates in real time.

Kraken
Kraken has an app available that provides real-time crypto market insights, including top movers, trending assets, and performance data. Here is a quick snapshot when I asked it to generate the market snapshot for today.

Connecting to Financial Datasets

In addition to the apps, I also tried connecting to Financial Datasets in two different ways. Financial Datasets is a developer friendly financial data provider that’s optimized for AI agents.

1. Custom GPT with External API Calls
You have more control with a custom GPT since you decide which actions to create. You can create actions for all available function calls. However, it requires more setup. I had to spend time working with ActionsGPT to ensure the schema was correct.

2. Customized App Connecting to an MCP Server
The setup is straightforward once you locate the integration URL. There is no need to define the schema, but you can only work with the functions provided by the MCP server.

How I Currently Use GPTs, Projects, and Apps

It has been exciting to experiment with GPTs, Projects, and Apps to increase productivity.

I use GPTs for generic tasks that can be repeated many times, such as my agent design brainstormer. I use Projects for specific research, as they consolidate multiple sources and chats in one place. For example, if I want to analyze the sustainability efforts of Google, I can create a project to iterate and collaborate if I have an enterprise plan. I use Apps to extend ChatGPT’s capabilities in different ways.

Besides querying financial data, for daily use you can also connect to the Target app for shopping or so.

That is what I have for this week. I will share more features about Claude next time. Have a great weekend.