Documentation

Extracto MCP server

Connect Claude, ChatGPT, Cursor or VS Code to Extracto and run scrapers from chat: browse the catalog, check the cost, run a task and read the results.

Overview

Extracto runs as a remote MCP (Model Context Protocol) server, so AI assistants such as Claude, ChatGPT, Cursor and VS Code can browse scrapers, estimate costs, run tasks and read results for you in chat.

Endpoint: https://extracto.cloud/api/mcp (Streamable HTTP). Every request is authorized with your Extracto account: either by signing in with OAuth, or with an API key.

Connect Claude

In claude.ai or Claude Desktop open Settings → Connectors → Add custom connector, paste the endpoint and save. Claude opens an Extracto window: sign in and click Allow.

text · Connect Claude
https://extracto.cloud/api/mcp

In Claude Code you can sign in the same way (run /mcp after adding the server) or pass an API key:

bash · Connect Claude
claude mcp add --transport http extracto https://extracto.cloud/api/mcp \
  --header "Authorization: Bearer ext_your_api_key"

Connect ChatGPT

Turn on developer mode for connectors in ChatGPT settings, create a connector with the endpoint above and choose OAuth authentication. ChatGPT opens the Extracto sign-in page; click Allow to finish.

Cursor, VS Code and other clients

Clients that accept custom headers can use an API key from the API keys page instead of OAuth:

JSON · Cursor, VS Code and other clients
{
  "mcpServers": {
    "extracto": {
      "url": "https://extracto.cloud/api/mcp",
      "headers": { "Authorization": "Bearer ext_your_api_key" }
    }
  }
}

Tools

Tool What it does
list_scrapers Catalog of scrapers with prices, filtered by keyword
get_scraper Input schema, example input, pricing and output fields of one scraper
estimate_cost Credits a run would reserve (free)
run_scraper Queue a run with a hard budget cap
get_run Status and progress; can wait up to 20 seconds
get_results Result items page by page
list_runs Your recent runs
cancel_run Stop a queued or running run
get_account Credit balance and API limits

A typical chat: “Find 30 coffee shops in Austin with phone numbers” → the assistant picks the Google Maps scraper, shows you the estimate, runs it after you agree on a budget, waits for the result and summarizes or tables it.

Spending and safety

  • Runs spend your credits (1,000 credits = $1), exactly like the REST API.
  • run_scraper requires max_budget_usd. The server refuses a run whose estimate is above it and never charges more than the cap.
  • Assistants are instructed to show the estimate and agree on the budget with you before running.
  • Runs above $4 also need a fresh estimate from the last 10 minutes, so a large run always follows an estimate you have seen.
  • Repeating the same run request within 10 minutes returns the existing run instead of starting a second one.
  • Signing in creates a key named after the app (for example “MCP: Claude”) on your API keys page. Revoke it there to disconnect the app at any time.

Troubleshooting

  • 401 / “Authentication required”: reconnect the app, or check that the API key is active.
  • “Not enough credits”: top up on the pricing page.
  • “Rate limit reached”: the assistant called too often; wait a minute. Limits are the same as for the API.
  • 10 active keys: an account can have up to 10 keys. Revoke an unused one before connecting another app.

See the API documentation for webhooks, idempotency and the REST endpoints the MCP tools use.