Finance MCP

finance-mcp

Live financial data for any LLM agent. Stock quotes · crypto · SEC filings · insider trades · XBRL financials · FX · macro

Why

Most finance APIs want a signup, a key, and a credit card before your agent can answer *"what's Apple trading at?"*

finance-mcp wraps five genuinely free public sources behind eleven MCP tools. No key for any of them. The only configuration is an email address, and only because the SEC insists on one.

The one you can't get anywhere else

Company insiders — officers, directors, 10% owners — must report every trade in their own stock to the SEC within two business days. It's public, it's structured, and essentially nobody reads it.

The catch is that most of it means nothing. Options vest. Shares get withheld to pay the tax on that vesting. None of it is a decision, and all of it gets reported as "insider selling."

get_insider_activity separates the two. Real output, Apple, one week in June:

OPEN-MARKET (a deliberate decision to trade)

Sold: 5 filing(s) 302.92K shares ≈ $87,566,268.55

MECHANICAL (vesting, option exercises, tax withholding — no decision): 1 filing(s)

Newstead Jennifer — SVP, GC and Secretary 2026-06-15 +30.10K @ — Option exercise / conversion 2026-06-15 −16.24K @ $296.42 Shares withheld for taxes ← not a trade

LEVINSON ARTHUR D — Director ★ 2026-05-06 −149.53K @ $284.57 Open-market sale ← a real decision ★ 2026-05-06 −100.47K @ $285.04 Open-market sale 2026-05-06 −5.00K @ — Gift

Same company, same month. One of those is a signal; the other is payroll. Every row links to the filing it came from.

It reports what was disclosed and does not interpret it. Insider buying and selling both have innocent explanations, and neither predicts the share price.

So we measured how often it matters

That claim is testable, so we tested it: every Form 4 filed by an insider at an S&P 100 company over twelve months — 2,969 filings — classified by whether the filing contains an actual decision to trade.

Roughly seven in ten insider filings contain no decision to trade at all. At the median company it's 83%.

A second result surprised us more. Corporations file Form 4s too, as 10% owners — and 53 filings, 1.8% of the sample, accounted for 65% of every disposed share. The largest single "insider" in the data was Honeywell International Inc, disposing of 317M shares of a company it was spinning off. Rank insiders by share count and the top of the list isn't people.

Method, caveats and the raw data are in analysis/. It re-runs from scratch with two commands.

Tools

How it fits together

flowchart LR

A["Claude Desktop Claude Code any MCP client"] -->|MCP over stdio| B["finance-mcp"] B --> C["Yahoo Finance equities · ETFs · indices "] B --> D["CoinGecko crypto "] B --> E["SEC EDGAR filings · XBRL · insider trades "] B --> F["ECB / Frankfurter FX "] B --> G["World Bank macro "]

The shared HTTP layer paces requests per host, retries 429/403/5xx with exponential backoff, honours Retry-After, and caches responses in memory — 30s for quotes, hours for filings and macro series.

Quickstart

Claude Code — no clone needed

claude mcp add finance -e SEC_USER_AGENT="Your Name your@email.com" -- npx -y finance-data-mcp

From source

git clone https://github.com/adididitagain/finance-mcp.git

cd finance-mcp npm install npm run build claude mcp add finance -e SEC_USER_AGENT="Your Name your@email.com" -- node "$(pwd)/dist/index.js"

</details>

Claude Desktop

Add to claude_desktop_config.json — on macOS, ~/Library/Application Support/Claude/claude_desktop_config.json — then restart the app:

{

"mcpServers": { "finance": { "command": "npx", "args": ["-y", "finance-data-mcp"], "env": { "SEC_USER_AGENT": "Your Name your@email.com" } } } }

To run from a local clone instead, use "command": "node" with "args": ["/absolute/path/to/finance-mcp/dist/index.js"].

Any other MCP client

The server speaks MCP over stdio. Run node dist/index.js and point your client at it.

Try it

What's Apple trading at?

Compare BTC and ETH over the last 24 hours.

Show me Microsoft's revenue for the last 5 years from their filings.

What 8-Ks has Tesla filed recently?

Which companies mention "quantum computing" in their 10-Ks?

What's India's GDP growth been over the past decade?

Convert 5000 USD to JPY.

Are Apple insiders actually selling, or is that just vesting?

Data sources

Configuration

[!IMPORTANT]

MCP clients do not pass your shell environment to the server. Set SEC_USER_AGENT in the client config shown above, not in .zshrc.

Development

npm run dev      # tsc --watch

npm test # 66 offline tests, stubbed fetch + real EDGAR fixtures, no network npm run smoke # drives all 11 tools against the live APIs

Notes and limitations

• Yahoo rate-limits by IP, and it cares what you claim to be. Sending a spoofed desktop-browser User-Agent gets you throttled hard — measured 0/8 successful requests with a Chrome UA versus 8/8 with an honest finance-mcp/0.1 one, alternating back to back. This server identifies itself honestly for that reason; don't "fix" it by pretending to be a browser. FX is served from the ECB instead, which has no rate limit.
• Quotes may be delayed up to ~15 minutes and are not exchange-official.
• SEC XBRL figures are as-filed. get_sec_financials labels each row by the period it covers, not the fiscal year of the filing it came from — EDGAR's own fy field describes the filing, so a 10-K restating a prior year would otherwise mislabel it.
• `search_sec_filings` ranks by EDGAR's relevance, which favours companies with your query in their *name*.
• World Bank data is annual and typically lags one to two years.
• Insider data is only as complete as the window you read. get_insider_activity reads the most recent limit Form 4s and reports how many it skipped — a large company files hundreds a year, so a small limit shows a recent slice, not a full picture. Form 4s are due within two business days of the trade, so there is a short reporting lag.

Disclaimer

An informational data tool. Nothing it returns is investment advice, and the data carries no accuracy or availability guarantee. Verify anything you intend to act on against an official source.

License

MIT © Aditya Bisht