How to Connect Claude to Live Financial Data for Stock Analysis
Ask Claude “what is Apple’s operating margin this year?” and you get one of two answers: a polite explanation that it has no access to live data, or a confident number from its training data that may be a year or more out of date. Language models are good at reasoning about numbers, but they are not a data source, and asked to compute a return on invested capital from memory they tend to produce plausible-sounding mistakes.
The fix is to give Claude a tool that fetches and computes the numbers for it. The Model Context Protocol (MCP) is the open standard Anthropic built for exactly this, and the Finance Toolkit MCP server is a free, hosted server that connects Claude to financial statements, 80+ ratios, valuation models, risk and performance metrics, technical indicators and macro data for 60+ countries. Every calculation runs through the open-source Finance Toolkit, so the number Claude reports is the number you would get by running the code yourself.
Free, hosted and open source
One URL gives Claude, ChatGPT, Cursor and other AI assistants access to 500+ financial methods, financial statements and macro data for 60+ countries.
What You Need
A Claude account
Custom connectors work in claude.ai and Claude Desktop on every plan; the free plan allows one, so this can be it. Claude Code supports MCP servers on every plan.
A Financial Modeling Prep key
Company data comes from Financial Modeling Prep. The free plan (250 requests a day, five years of US data) is enough for this article. Macro data needs no key.
You enter the key once, on a standard OAuth page the server opens on first use. It is passed along with each request and never stored. If you would rather not use the hosted server, the local installation runs the same server on your own machine with one uvx command.
Connecting Claude Desktop or claude.ai
The steps are identical for the desktop app and the web app:
- Click Customize in the left sidebar and open the Connectors tab.
- Click + and select Add custom connector.
-
Enter a name, e.g. Finance Toolkit, and paste the server URL:
https://financetoolkit.jeroenbouma.com/mcp - Click Add. The Finance Toolkit now appears in your list of connectors.
- Start a new chat and ask a financial question. On the first tool call a browser window asks for your FMP key; enter it once and you are done.
There is nothing to install and nothing to keep up to date: the hosted server always runs the latest Finance Toolkit release.
Connecting Claude Code
In the terminal, one command registers the server for every future session:
claude mcp add --transport http finance-toolkit https://financetoolkit.jeroenbouma.com/mcp
Restart Claude Code and the tools are available; the first tool call triggers the same one-time key prompt. This combination works well for building on the numbers: Claude Code pulls the data through the server and writes the Python, notebook or report around it.
What You Can Ask
You never have to name a tool or an indicator: ask in plain English and Claude picks the right one. Below are ten real questions, each answered through the server with data up to October 2, 2026. The question is what you type, the chart is what comes back. Paste any of them into Claude as they are, or swap in your own tickers.
Which is more profitable, Apple or Microsoft?
Compare the gross, operating and net margins of Apple and Microsoft over the last five years. Which one is more profitable?Finance Toolkit MCP · profitability
On margins the answer is Microsoft, by a wide and steady gap. Without being asked, Claude then checks capital efficiency, and that points the other way: Apple’s return on invested capital was 70.38% in 2025 against 30.64% for Microsoft.
Microsoft earns more per dollar of revenue; Apple earns more per dollar of capital. A single headline number would hide that.
What drives the difference between Coca-Cola and PepsiCo?
Run a DuPont analysis on Coca-Cola and PepsiCo and explain what drives the difference in return on equity.Finance Toolkit MCP · models · DuPont analysis
Coca-Cola mostly sells concentrate and leaves the capital-heavy bottling to partners, so it keeps 27 cents of every dollar as profit. PepsiCo owns its snack factories and distribution: thinner margins, but it turns its assets over twice as fast and uses more leverage (an equity multiplier of 5.3 versus 3.4). Switch between the tabs to see each component.
A similar return on equity, built in opposite ways: Coca-Cola through margin, PepsiCo through volume and leverage.
What growth is priced into Nvidia?
What is the intrinsic value of Nvidia based on a discounted cash flow, and how does it compare to the current share price?Finance Toolkit MCP · models · intrinsic valuation
A DCF is only as good as its growth assumption, so Claude runs it at four growth rates (10% cost of capital, 3% terminal growth, five years):
The share price of $233.95 only fits with roughly 40% annual cash flow growth for five years. The question is whether you believe that, not whether the stock is cheap.
Which carmaker is in the weakest financial position?
Calculate the Altman Z-Score and Piotroski F-Score for Ford, General Motors and Stellantis.Finance Toolkit MCP · models · Altman Z-Score, Piotroski F-Score
All three ended 2025 below 1.8, the classic distress threshold. Stellantis moved the most, from 2.27 in 2023 as its operating result turned negative. A good answer adds the caveat: Ford and GM run large car-finance arms whose loan books inflate liabilities and push the score down structurally.
Read the trend within each company, not the gap between them. Stellantis is the one that deteriorated.
Is the semiconductor sector overbought?
Is the semiconductor sector overbought? Look at the RSI for NVDA, AMD, TSM and ASML.Finance Toolkit MCP · momentum · RSI
AMD was the outlier of the year, above 70 for most of April to July with a peak of 84.6 in May. Nvidia never reached overbought territory in 2026 and touched 38.3 in March.
AMD and TSMC are back above 70; Nvidia and ASML are not. An RSI above 70 says the move was strong, not that it must reverse.
Did Tesla have a golden cross or a death cross?
Has Tesla had a golden cross or a death cross this year based on the 50-day and 200-day moving averages?Finance Toolkit MCP · overlap · moving averages
The 50-day average fell below the 200-day on April 9, 2026, and there has been no golden cross since. The gap kept widening after the July drop to around $311; the price has recovered to about $371, above the short average but still below the long one.
A death cross in April, and the trend has not turned since.
How did Amazon, Alphabet and Meta do on a risk-adjusted basis?
Compute the Sharpe ratio, Sortino ratio, alpha and beta of Amazon, Alphabet and Meta against the S&P 500 since 2020.Finance Toolkit MCP · performance
2022 was bad for all three on every measure. 2023 belonged to Meta after its “year of efficiency”, 2025 to Alphabet, with the market’s best alpha of the three and the lowest beta. The Sharpe and Sortino ratios are calculated within each year, so compare them with each other rather than with annualised figures elsewhere.
No single winner: leadership rotated from Meta in 2023 to Alphabet in 2025.
How much downside risk do Berkshire Hathaway, Visa and Costco carry?
What is the 95% Value at Risk and the maximum drawdown of Berkshire Hathaway, Visa and Costco since 2020?Finance Toolkit MCP · risk · VaR, CVaR, maximum drawdown
Berkshire has the calmest days, with a daily VaR of −1.83%, close to the S&P 500. Visa has the fattest daily tail (CVaR −3.93%). But Costco had the smallest worst-case fall since 2020 (−20.3%), while Berkshire’s −24.3% was slightly deeper than the market’s.
Daily risk and drawdown risk rank the same stocks differently. Ask for both.
How have unemployment rates developed since 2010?
Show me the unemployment rate for the United States, Germany and Japan since 2010.Finance Toolkit MCP · jobs · unemployment rate
The US went from 9.6% in 2010 to 3.7% in 2019, jumped to 8.1% in 2020 and was back at 3.6% two years later; it has loosened to 4.4% since. Japan has been flat at about 2.5% for years.
Germany barely moved in 2020 (3.6%): its short-time work scheme kept people employed while hours were cut.
How do inflation and interest rates compare in the US and the Eurozone?
Compare inflation and central bank interest rates in the Eurozone and the United States over the last few years.Finance Toolkit MCP · macroeconomics, rates
The Federal Reserve raised rates earlier and further. The real policy rate, the policy rate minus inflation in the third tab, shows the difference best. Policy rates are annual averages.
US policy was restrictive in real terms from 2023; the Eurozone only from 2024, after a real rate of −2.3% in 2023.
Every answer maps to a documented Finance Toolkit method, so you can look up exactly how a number was calculated in the documentation. The MCP server page has the full conversations for six more questions.
Ask your own questions
These ten examples use a fraction of what the server can do. Connect it in a minute and ask about any company, index, sector or economy: the answers are calculated, not recalled from memory.
Tips for Better Answers
Name the period
"Since 2020" avoids Claude picking a range for you. Company data follows fiscal years: September for Apple, June for Microsoft.
Ask for the definition
P/E and ROIC have several definitions. Ask "which formula was used?"; every one is documented.
Let it chain
The best answers come from several tool calls in a row: margins, then returns, then valuation. If it stops early, ask it to continue.
Pick a capable model
Every model gets the same data; the depth of the interpretation scales with the model.
Mind the free plan
FMP's free plan allows 250 requests a day. A broad screen can use that quickly, so start with a few companies.
Build on it
In Claude Code, ask for the notebook or script around the answer; the numbers stay the server's.
Not a Claude User?
The same server works with any MCP-compatible client, each with step-by-step instructions on the Finance Toolkit MCP server page:
Clients that only support the stdio transport can use the local installation. If you would rather skip the assistant altogether, the Finance Toolkit Python library exposes the same 500+ methods directly. Problems or ideas for a new tool are welcome as an issue on GitHub.