Extracto

Model Context Protocol server for Extracto. It gives Claude, Cursor, Claude Code, and any MCP client the ability to turn a URL plus a schema into validated, typed JSON — no prompt engineering, no HTML parsing, and no hallucinated fields (missing data comes back as null).

Quick start

You need an Extracto API key. Get one at app.getextracto.dev/keys.

The server runs over stdio and is published to npm, so most clients just need this config block.

Claude Desktop

Edit claude_desktop_config.json (Settings → Developer → Edit Config):

{
  "mcpServers": {
    "extracto": {
      "command": "npx",
      "args": ["-y", "extracto-mcp"],
      "env": { "EXTRACTO_API_KEY": "exa_live_your_key_here" }
    }
  }
}

Cursor

Add to ~/.cursor/mcp.json (or the project .cursor/mcp.json) with the same block.

Claude Code

claude mcp add extracto -e EXTRACTO_API_KEY=exa_live_your_key_here -- npx -y extracto-mcp

Restart the client and ask it to extract something, e.g. "Use extracto to pull the title, language and star count from github.com/facebook/react."

Tools

Tool	What it does
extract	Synchronous extraction from a single URL (up to ~90s). Returns { data, meta }.
extract_async	Submit an async job for heavy or anti-bot pages. Returns a job id immediately.
get_job	Poll an async job for status and result.
list_jobs	List your recent async jobs.

The schema argument

A schema is an object mapping field names to types. A type is:

• a literal: "string", "number", "boolean", "array", "object"
• a one-element array for a list: ["string"], or [{ "title": "string" }]
• a nested object: { "author": { "name": "string" } }

{
  "title": "string",
  "price": "number",
  "tags": ["string"],
  "reviews": [{ "user": "string", "stars": "number" }]
}

Only fields that are actually found on the page are returned; anything missing is null rather than guessed.

Configuration

All configuration is via environment variables passed by your MCP client:

Variable	Required	Description
EXTRACTO_API_KEY	yes	Your key from app.getextracto.dev/keys.
EXTRACTO_BASE_URL	no	Override the API host (defaults to https://app.getextracto.dev).
EXTRACTO_TIMEOUT_MS	no	Per-request timeout in ms (default 90000).

Development

npm install
npm run dev        # run from source with tsx
npm run typecheck
npm run build      # bundle to dist/ with tsup

Related

• extracto — the official TypeScript/JavaScript SDK.
• getextracto.dev — docs and API reference.

License

MIT