Ticker Scout MCP server

An agent-native data layer for SEC filings, over the Model Context Protocol. Finished financial statements for public companies, and the filings themselves read and synthesized. The SEC publishes the numbers as data, but it does not publish them as statements. This serves them already assembled, through 20 tools: any statement for any fiscal year or quarter, any line item by its standard name or the company's own, every fiscal year and quarter held, every accession number cited, and the 10-K, 10-Q and 8-K narrative by section, topic or keyword. No API key, no signup, no paywall.

https://mcp.tickerscout.ai/mcp

It exposes the data published at tickerscout.ai (https://tickerscout.ai): financial statements, the annual and quarterly reports read and synthesized, and material event histories for S&P 500 companies and major foreign issuers that file with the SEC, all derived from their own filings.

What is already done for you

The difference between this and the EDGAR APIs is assembly, not availability. The SEC gives you facts. This gives you statements.

• The three statements arrive as statements. Income statement, balance sheet and cash flow, line items in reported order, not a flat bag of tagged facts you have to select concepts for and group yourself.
• Assembled across filings, because no single filing holds the whole record. Annual figures come from the annual report, quarterly and year-to-date figures from the quarterly report, and where a quarterly report presents cash flow only on a year-to-date basis, the discrete quarter comes from the earnings release filed as a current-report exhibit. For a US filer that means the 10-K, the 10-Q and the 8-K; for a foreign issuer, the 20-F or 40-F and the 6-K.
• On the company's own fiscal calendar. 52- and 53-week years, quarters that are 13-week periods rather than calendar quarters, and fiscal years that do not end in December. Period dates come from the column headings the filings themselves print.
• In actual dollars and actual shares. Never thousands, never millions, no scale factor to infer. Share counts are adjusted for splits. A foreign issuer's figures are converted to US dollars at the rate recorded on each period, and per-share figures are per ADS where the company has an ADR program.
• Shaped to the business. A bank, an insurer, an asset manager and a retailer do not have the same statements, so they do not get the same sections. Ask what a company has rather than assuming a fixed shape.
• With the gaps named. Where a filing does not disclose a figure it is omitted and the omission is explained, never silently zeroed. Every figure cites the SEC accession number it came from.
• Figures link to the line of the filing that prints them. Wherever the filing tags a figure, the response carries a sec.gov link that opens the filing at that number.

Connect

Claude Code

claude mcp add --transport http ticker-scout https://mcp.tickerscout.ai/mcp

Claude on the web or desktop

Settings, then Connectors, then Add custom connector, and paste the URL.

Cursor, and any client with an mcpServers config

{
  "mcpServers": {
    "ticker-scout": {
      "url": "https://mcp.tickerscout.ai/mcp"
    }
  }
}

Clients that only speak stdio

{
  "mcpServers": {
    "ticker-scout": {
      "command": "npx",
      "args": ["mcp-remote", "https://mcp.tickerscout.ai/mcp"]
    }
  }
}

Start here

The main six tools answer most questions in one call; the other fourteen each answer one kind of question, and take arguments rather than multiplying into near-duplicates.

Tickers are case-insensitive, and BRK.B and BRK-B both resolve.

How the data is organized

Records and periods

A record is the complete snapshot built when a company filed one 10-Q or 10-K, tagged by fiscal period: FY27Q2 is the record built from NVIDIA's second-quarter fiscal 2027 10-Q, FY26 one built from a 10-K. Each record holds several periods: up to three fiscal years, the latest quarter and its prior-year comparative, year to date and trailing twelve months. Older records hold older quarters, so a company's history is the union of its records, and every record stays published at a permanent dated URL.

Tools that take period search every record automatically. Tools that take record read one snapshot, the latest by default.

The period argument

A period earlier than anything held is not an error. The answer says the data does not go back that far and lists exactly what is held. A period not yet reported says so and gives the next expected filing.

Names in the record

Records are shaped to each business, so the same line item has different names: Apple's revenue is total_net_sales, JPMorgan's is total_net_revenue, Walmart's net income is net_income_attributable_to_walmart. Nothing is relabeled to hide that.

• Every value a figure tool returns leads with name_in_record (the record's own name), field (its path inside the period) and json_path (its exact address in financials.json, e.g. annual[0].income_statement.total_net_sales).
• Any of those names can be passed back to get_figure exactly as written, and returns exactly that field. Case, spaces and underscores do not matter: Data Center and revenue_by_market_platform.Data Center both work.
• A standard figure asked for by its generic words (revenue, net income, free cash flow) resolves the standard way below, each value still under the record's own name. If the record also has other fields of that name (NVIDIA's segment tables each carry a revenue), they are listed under other_fields_with_this_name, callable by path.
• list_company_figures maps each standard figure to the name this company uses, then lists every remaining figure under its own name. Pass query to search those names.

Standard figures

get_figure and compare_companies know the names each figure goes by across companies and resolve deterministically: first the field the company's own headline uses for it, then a list of known names in order, then a name pattern only when exactly one field matches. Nothing is estimated, derived or summed to fill a gap. Look-alike fields in the same statement come back under related, and a company that reports none of the names gets a plain "not reported" with the closest fields it does have.

guide() returns the full catalog with every word each figure answers to.

Units and sources

All money is in actual dollars (not thousands, not millions), per-share figures in dollars per share, share counts in actual shares, split-adjusted. Every figure response carries the record's own units statement, so a value is never separated from its scale.

Every value carries the SEC accession number of the filing it came from and, where the filing tags the number, a sec_link that opens the filing at the line printing it. A value with no link was calculated in the record (margins, free cash flow, trailing twelve months, a quarter derived by subtraction) or comes from an earnings release, which the SEC does not tag. Margins and ratios this server calculates (get_margins_and_ratios) are labeled calculated and show their formula and the exact fields used.

All 20 tools

Every tool is read-only. Tickers are case-insensitive. A period argument takes the syntax above and searches every record; a record argument reads one earlier snapshot.

The main six

Statements and periods

Figures

Search, filings and conventions

Examples

> What did NVIDIA earn last quarter?

  get_key_figures(ticker: "NVDA")

> Apple's revenue for fiscal 2025, and where the 10-K prints it

  get_figure(ticker: "AAPL", figure: "revenue", period: "FY2025")

  name_in_record  total_net_sales
  json_path       annual[0].income_statement.total_net_sales
  value           416161000000   ($416.16 billion)
  accession       0000320193-25-000079
  sec_link        https://www.sec.gov/Archives/edgar/data/320193/000032019325000079/aapl-20250927.htm#f-78

> NVIDIA revenue in fiscal 2019

  get_figure(ticker: "NVDA", figure: "revenue", period: "FY2019")

  available: false
  We don't have data going back that far: FY2019 is earlier than anything Ticker Scout
  holds for NVDA. The earliest fiscal year held is FY2024. ...

> Microsoft's revenue, net income and EPS in one call

  get_figure(ticker: "MSFT", figure: ["revenue", "net income", "eps"], period: "latest quarter")

> NVIDIA Data Center revenue, every quarter held

  get_figure(ticker: "NVDA", figure: "Data Center", period: "all")

Data policy

• Everything is derived from public reports filed with the SEC, such as annual reports on Form 10-K, quarterly reports on Form 10-Q and current reports on Form 8-K (20-F, 40-F and 6-K for foreign issuers). Nothing is estimated and nothing is fabricated. Where a filing does not disclose a figure, it is omitted and the omission is explained.
• No market price data. No quotes, no market caps, no price snapshots. This is a fundamentals source.
• Data is refreshed each quarter, shortly after a company files. Coverage and the period held for each company are in list_companies.
• This is factual synthesis of public filings. It is not investment advice.

Free to read and cite. Please attribute "Ticker Scout (tickerscout.ai)".

How it works

A stateless Cloudflare Worker using createMcpHandler from @modelcontextprotocol/server, speaking streamable HTTP at /mcp.

The server stores no data. Every tool call fetches the corresponding files from https://tickerscout.ai through the Cloudflare edge cache and shapes the response; a question about an older period reads that record's permanent dated URL. There is no database, no snapshot and no fallback copy: if an upstream file cannot be fetched, the tool returns an error naming the URL and the status rather than serving something stale.

src/
  index.ts       Worker entrypoint (exports the handler only)
  server.ts      Builds the tool definitions once per isolate, a server per request
  upstream.ts    Fetching, edge caching, the error policy
  tickers.ts     Ticker normalization, coverage search, near matches
  records.ts     A company's published records, and choosing which to read
  periods.ts     Fiscal periods: naming, the period syntax, out-of-range answers
  figures.ts     Standard figures, name resolution, figure_sources links
  filings.ts     Accession numbers, EDGAR links, citations in prose
  topics.ts      Narrative and events topics, keyword search
  markdown.ts    Section parsing for the .md documents
  financials.ts  Headline condensing and unit-safe slicing
  respond.ts     Result shaping, source URL and attribution
  tools/         Tool registration, one module per group

Development

npm install
npm test                      # unit tests for the pure modules
npm run type-check            # wrangler deploy does not typecheck
npm run dev                   # local server on http://localhost:8787/mcp
npm run smoke -- <url>        # live end-to-end check against a deployment
npm run sweep -- <site dir>   # every tool, every company, against a local copy of the site
npm run deploy

npm test covers the pure logic. test/smoke.mjs calls the tools against a running server for a deliberately shape-diverse set of companies (a conventional company, a bank, an insurance conglomerate, a retailer far ahead of the calendar, a foreign issuer) plus negative cases. test/sweep.mjs runs every tool for every covered company in-process, plus every topic, statement, breakdown and metric kind, serving the site's files from a local directory, and fails on any error. A unit test keeps tools/list under 40 KB.

License

MIT. See LICENSE.