Web data. Browser actions.
One MCP connection.

Choose how to connect

Use the hosted endpoint if your client supports remote MCP. Prefer a local process? Use the npm or Python package with your API key. You need a FineData account either way.

{
  "mcpServers": {
    "finedata": {
      "command": "npx",
      "args": ["-y", "@finedata/mcp-server"],
      "env": { "FINEDATA_API_KEY": "fd_your_api_key" }
    }
  }
}

{
  "mcpServers": {
    "finedata": {
      "url": "https://mcp.finedata.ai/mcp",
      "headers": { "Authorization": "Bearer fd_your_api_key" }
    }
  }
}

{
  "mcpServers": {
    "finedata": {
      "url": "https://mcp.finedata.ai/mcp"
    }
  }
}

Make one call before building a loop

Reload your client's MCP connection and check that nine tools appear. Ask your agent:

Use FineData to read https://example.com as markdown. Report the page title and the tokens used. Do not enable extra rendering modes.

{
  "url": "https://example.com",
  "formats": ["markdown"],
  "only_main_content": true
}

More than page-to-markdown

Use the same connection for structured extraction, multi-page research and browser interactions. These are tool arguments, not client configuration.

{
  "url": "https://example.com/product",
  "extract_schema": {
    "type": "object",
    "properties": {
      "name": {
        "type": "string"
      },
      "price": {
        "type": "number"
      },
      "currency": {
        "type": "string"
      }
    },
    "required": [
      "name",
      "price",
      "currency"
    ]
  }
}

{
  "urls": [
    "https://example.com",
    "https://example.org"
  ],
  "formats": [
    "markdown"
  ],
  "only_main_content": true
}

{
  "url": "https://example.com/catalog",
  "use_js_render": true,
  "js_wait_for": "selector:.product",
  "formats": [
    "markdown"
  ]
}

{
  "url": "https://your-api.example/items",
  "method": "POST",
  "headers": {
    "Content-Type": "application/json"
  },
  "body": "{\"name\":\"Example item\"}",
  "auto_retry": false,
  "formats": [
    "text"
  ]
}

For slow pages, use scrape_async, save its job_id, then poll get_job_status with a delay. Stop on completed, failed or cancelled. MCP async and batch tools are GET-only; writes use the synchronous send_http_request tool.

The nine tools

Tool annotations describe read-only, destructive and idempotent behavior; they are hints for the client, not permission checks or a budget cap. scrape_url sends GET; send_http_request handles HTTP writes. Browser actions such as clicking or typing can also change a site, even during a GET-based scrape. Review those actions and your client's confirmation settings.

Use list_jobs to resume asynchronous work, not to retrieve synchronous scrape history. Only pending jobs can be cancelled; processing jobs finish normally. For complete field definitions, see the API options and your client's tool schema.

Sign-in instead of a shared key

An API key is a bearer secret: whoever holds it is you. That is fine for an agent you run yourself and awkward for one you hand to a teammate or a customer. The remote endpoint therefore speaks OAuth 2.1, so each person connects their own account.

Why cost matters more for agents

If the first call fails

No tools appear

Check the client's MCP logs. For local setup, verify Node.js/npm is available to the app and the key is in env. Reload the connection. For remote setup, use the full URL ending in /mcp and a client with Streamable HTTP support.

401 or OAuth sign-in failure

Verify the key or reconnect with OAuth. A remote API key belongs in Authorization: Bearer fd_…, not in the URL. Keep keys out of shared repositories and screenshots.

402, 403 or 429

Read the error: check billing for 402, account or feature access for 403, and reduce request frequency for 429. Changing the rendering engine does not fix these errors.

A job is still running

Keep the returned job_id or batch_id and poll that job with a delay. Do not submit the same work again. A batch can finish with partial failures; inspect individual results.

Questions

Related Guides