GetABrain MCP server

@getabrain/mcp-server

MCP server for GetABrain.ai — give your AI agent real human judgment as native tools.

Use with Claude Desktop / Cursor

Add to your MCP client config (e.g. claudedesktopconfig.json):

{
  "mcpServers": {
    "getabrain": {
      "command": "npx",
      "args": ["-y", "@getabrain/mcp-server"],
      "env": {
        "GETABRAIN_API_KEY": "gab_k_…",
        "GETABRAIN_API_SECRET": "gab_s_…"
      }
    }
  }
}

Get your API key by signing up at https://getabrain.ai.

Remote (hosted) MCP server -- no install

Prefer not to run anything locally? GetABrain also hosts this same server over Streamable HTTP at https://www.getabrain.ai/api/mcp. Point any MCP client that supports remote servers at that URL and pass your key pair as headers instead of env vars:

{
  "mcpServers": {
    "getabrain": {
      "url": "https://www.getabrain.ai/api/mcp",
      "headers": {
        "X-API-Key": "gab_k_…",
        "X-API-Secret": "gab_s_…"
      }
    }
  }
}

Same 7 tools, same schemas, same test-mode support -- see docs/deploy/remote-mcp.md in this repo for details (Smithery-style clients, auth requirements, etc).

Test mode

Test mode is a flag on the key, not a different key format. When you mint an API key — via POST /api/v1/requestor/keys with {"mode":"test"}, or by choosing "test" in the dashboard — you get back a completely normal gabk… / gabs… key pair. There's no test in the string; the test-ness lives in the database as an is_test flag on that key. No funding or card required.

Point GETABRAINAPIKEY / GETABRAINAPISECRET at a test-mode key and the server behaves identically, except:

wait/poll → rate) can be built and exercised end-to-end before any real human worker or real money is involved.

• submitquery never touches your balance — no charge, no insufficientbalance errors.
• Responses come back synthetic and are always marked simulated: true, so your pipeline (submit →

• get_balance reports mode: "test" so the agent/human can tell at a glance which environment it's in.

When you're ready to go live: mint a live-mode key (same call, {"mode":"live"} or the dashboard default), fund the account with createtopuplink (works with either key type — a test-mode agent can generate the link, a human completes checkout to add real funds), and swap the env vars. getbalance then reports mode: "live", and submitquery starts spending real balance and dispatching to real paid workers.

Tools

(with a setup link + hint when it's off and would otherwise stall a live account at zero balance).

• getbalance — read-only: prepaid balance (cents), mode ("test"/"live"), and autoreload_enabled

pay — the agent cannot complete payment itself.

• createtopuplink — mints a Stripe Checkout URL to add funds (min $5); a human opens it in a browser to

image/video/audio review, voice/video/photo capture, …). Returns a query_id. Spends balance on a live key; free and simulated: true on a test key.

• submit_query — ask real humans a question (16 query types: A/B test, rating, ranking, sentiment, yes/no,

responses once enough arrive, or pending — call again to keep waiting. Use this instead of get_responses when you want the tool call itself to wait.

• get_responses — one-shot, read-only: current status + whatever responses exist right now, no waiting.
• waitforresponses — bounded polling (up to maxwaitseconds, default/max 50s); returns ready with

• list_queries — read-only: your recent queries, optionally filtered by status.
• rateresponse — rate a worker's answer 1–5 (optional feedbacktext); feeds the worker quality system.

Example agent flow

Shortened. The full README is on GitHub.