MCP server

The Belfort Data MCP server gives an AI agent the API’s routes as tools. It runs on your machine over stdio, calls the public API with your key, and never touches a database. Every tool is read-only.

Not on npm yet

The server isn’t published as a package yet. The steps below run it from a copy of the Belfort Data API source, which needs Node.js 22 or later and pnpm. During the pilot, ask Belfort for access.

1. Install

In your copy of the source:

pnpm install

The server reads two environment variables:

Variable	Needed	Default
BELFORT_API_KEY	For every tool except status
BELFORT_API_URL	No	https://belfort-data-api.vercel.app. Must be https, except for localhost.

The key is sent as X-API-Key. The server never logs it and never puts it in a tool’s output.

2. Add it to your client

Replace /absolute/path/to/belfort-data-api with where your copy lives.

Claude Code

claude mcp add --transport stdio belfort-data \
  --env BELFORT_API_KEY="$BELFORT_API_KEY" \
  -- pnpm --dir /absolute/path/to/belfort-data-api exec tsx mcp/index.ts

Then run claude mcp list to check it connected.

Claude Desktop

Open Claude Desktop’s settings, choose Developer, then Edit Config, and add the server to claude_desktop_config.json:

{
  "mcpServers": {
    "belfort-data": {
      "command": "pnpm",
      "args": ["--dir", "/absolute/path/to/belfort-data-api", "exec", "tsx", "mcp/index.ts"],
      "env": {
        "BELFORT_API_KEY": "bda_..."
      }
    }
  }
}

Restart Claude Desktop. The tools appear under the server’s name.

Other clients

Any MCP client that runs a local stdio server can use it. Run this command with BELFORT_API_KEY set in its environment:

pnpm --dir /absolute/path/to/belfort-data-api exec tsx mcp/index.ts

Tools

Tool	Route	Key	Inputs
status	GET /v1/status	No	None
get_company	GET /v1/companies/{id}	Yes	id
get_metrics	GET /v1/companies/{id}/metrics	Yes	id, optional metric and frequency
get_statements	GET /v1/companies/{id}/statements	Yes	id, optional statement, form, limit, cursor
get_financials	GET /v1/companies/{id}/financials	Yes	id, optional field, period, limit, cursor
get_segments	GET /v1/companies/{id}/segments	Yes	id, optional axis, period_end, limit, cursor
get_insider_trades	GET /v1/companies/{id}/insider-trades	Yes	id, optional limit, cursor
get_form13f	GET /v1/companies/{id}/form13f	Yes	id (a manager’s CIK), optional limit, cursor
get_holders	GET /v1/companies/{id}/holders	Yes	id, optional period, limit, cursor
get_ownership	GET /v1/companies/{id}/ownership	Yes	id, optional form, limit
get_filings	GET /v1/companies/{id}/filings	Yes	id, optional limit, cursor
list_earnings_releases	GET /v1/companies/{id}/earnings-releases	Yes	id, optional limit, cursor, from, to
get_earnings_release	GET /v1/companies/{id}/earnings-releases/{accession}	Yes	id, accession
get_macro	GET /v1/macro/{series}	Yes	series

The inputs follow the API’s parameters exactly; see each dataset for the values. An unknown or invalid input is refused before any API call.

What to expect

• One call, one page. A tool returns one page of the API’s answer. To read further, call it again with meta.next_cursor as cursor. A default statements page of four documents is about 45 KB, so ask for what you need.
• Errors come back as tool errors, with the API’s code and message, including Retry-After on a 429 or 503.
• The caveats travel with the tools. Each tool’s description tells the model that a figure can be withheld, that it can change when a later filing arrives, that adjustments explain what we changed, and that this is data, not investment advice.
• No caching and no guessing. The server adds nothing to the API’s answers. Its rate limits and data gaps are the API’s: see Requests and responses and Known limits.