edgar-mcp

An MCP server that lets Claude search the full text of SEC filings.

Ask Claude "which companies discussed agentic AI in their 10-Ks this year?" and it searches EDGAR directly — 235 real filings, with links.

The hardest problem

EDGAR filters by CIK, a ten-digit identifier that means nothing to a human and everything to the API. The obvious design accepts a CIK as a tool parameter and lets the model supply it. That fails in the worst possible way: a language model will confidently invent a CIK for any company it hasn't memorized, and a wrong company's filings look exactly like the right company's — correctly formatted, plausible, and completely wrong, with nothing to signal the error. So the tools accept a ticker or company name and resolve it in code against the SEC's own mapping file. Where the name is ambiguous — "American" matches 66 companies — the resolver returns the candidate list rather than picking the best-scoring match, and Claude asks which one you meant. The model handles the easy cases through context; the code catches the ones it gets wrong.

Install

npm install -g @alinarashid/edgar-mcp

Add to claude_desktop_config.json:

{
  "mcpServers": {
    "edgar": {
      "command": "npx",
      "args": ["-y", "@alinarashid/edgar-mcp"],
      "env": {
        "SEC_USER_AGENT": "your-app [email protected]"
      }
    }
  }
}

SEC_USER_AGENT is required. The SEC rejects requests that don't identify the caller, and a missing header returns 403 rather than an empty result.

Restart Claude Desktop.

Tools

search_filings — full-text search across every filing since 2001. Filter by form type, date range, and company. Returns the filing company, form, date, and a document link.

list_company_filings — one company's filings, newest first. Accepts a ticker or name. Filter by form type, or you will mostly get Form 4 insider trades.

Examples

Which companies mentioned "agentic AI" in their 10-K filings this year?

Show me American Airlines' recent quarterly reports.

Find 8-K filings from 2026 that mention "material weakness".

Known limits

• Search returns metadata, not text. It tells you which filings match, not what they say. Follow the URL to read the language.
• Ranking is keyword relevance, not company size. A short filing that repeats a phrase outranks a 300-page 10-K that mentions it twice.
• Public companies only. Private companies file nothing with the SEC.
• Searching within a company includes third-party filings tagged with that CIK, such as shareholder proposals filed by outside groups.

Notes

The full-text endpoint at efts.sec.gov is undocumented — the SEC publishes no parameter list, response schema, or stability guarantee. This package reads defensively and may need updating if the shape changes.

Requests are serialized to roughly 8 per second, under the SEC's published ceiling of 10.

License

MIT