GetABrain MCP Server is an MCP server for GetABrain.ai, whose core goal is to invoke real human judgment as a native tool for AI agents.
Purpose Enable AI agents, when human judgment is needed, to ask paid human workers questions through MCP tools and receive structured, schema-compliant JSON answers. Use cases include A/B testing, rating, ranking, sentiment analysis, yes/no judgments, image/video/audio moderation, and voice/video/photo capture, among others.
Main capabilities The server provides 7 tools:
- get_balance: read-only, returns the prepaid balance (in cents), mode (test/live), and auto_reload_enabled; when auto-reload is off and the live account balance may reach zero, it includes a setup link and a prompt.
- create_topup_link: generates a Stripe Checkout URL for topping up (minimum $5), which requires a human to open a browser and complete payment; an agent cannot complete payment on its own.
- submit_query: asks a human a question, supports 16 query types (A/B testing, rating, ranking, sentiment, yes/no, image/video/audio moderation, voice/video/photo capture, etc.), and returns a query_id. Live keys consume balance; test keys are free and return simulated: true.
- get_responses: one-shot, read-only, immediately returns the current status and existing responses without waiting.
- wait_for_responses: bounded polling, up to max_wait_seconds (default and maximum both 50 seconds); when enough responses have arrived, returns ready and the responses; otherwise returns pending, and can be called again to continue waiting.
- list_queries: read-only, lists recent queries, and can be filtered by status.
- rate_response: rates a worker's answer on a 1–5 scale (optional feedback text), feeding back into the worker quality system.
How to use Local use: add the server to the MCP client configuration (for example, claude_desktop_config.json), and in https://getabrain.ai register to obtain an API key, and pass it in via the environment variables GETABRAIN_API_KEY / GETABRAIN_API_SECRET. Remote hosting: no local installation is required; GetABrain uses Streamable HTTP at https://www.getabrain.ai/api/mcp Serve the same server; point MCP clients that support remote servers at this URL and pass the key pair in a request header instead of an environment variable. The remote version provides the same 7 tools, the same schema, and test mode support; for details about Smithery-style clients, authentication requirements, and more, see docs/deploy/remote-mcp.md in the repository. Test mode: Test mode is a flag on the key, not a different key format. By POSTing /api/v1/requestor/keys with {"mode":"test"}, or selecting "test" in the dashboard, you get a completely normal gab_k_… / gab_s_… key pair; there is no test in the string, and test status exists in the database as an is_test flag. No top-up or bank card is required. After pointing the environment variables at the test key, the server behaves consistently, but submit_query does not touch the balance, there is no charge, and there are no insufficient_balance errors; responses are synthetic data and are always marked simulated: true, making it easy to build and rehearse the submit → wait/poll → rate pipeline end to end before involving real people or real money; get_balance reports mode: "test". Switching to live: Generate a live-mode key (same call, passing {"mode":"live"} or using the dashboard default), top up the account with create_topup_link (both key types can generate links; a test-mode agent can generate a link, and a real person completes checkout to add real funds), then replace the environment variables. At this point get_balance reports mode: "live", and submit_query starts consuming the real balance and dispatching to paid human workers.
Example agent flow
- get_balance confirms funds (or mode: "test" for a free sandbox run).
- If the live key has insufficient funds: create_topup_link → a real person completes checkout → get_balance again.
- submit_query → obtain a query_id.
- wait_for_responses (repeat while pending) → read the human (simulated in test mode) answers.
- rate_response → optionally for each respScore responses to improve future worker matching.
Limits Queries wait up to 50 seconds; after a timeout, call again. Payment must be completed by a real person; agents cannot pay on their own. Test mode returns synthetic data and does not represent real human results. For complete API documentation, see https://getabrain.ai/docs/api。
