The Model Context Protocol gave agents a clean way to reach tools — but almost every MCP server on the registry wraps software: a database, a search index, a SaaS API. GetABrain is an MCP server that wraps people. When your agent calls the submit_query tool, a real human answers, the answer comes back quality-scored, and the agent keeps going. This is the human-in-the-loop step as a native MCP tool call, not a Slack channel someone has to babysit.
Why a human-in-the-loop MCP server at all
Agents built on MCP hit the same wall every agent hits: some steps a model genuinely should not decide alone. Is this generated marketing claim actually true? Which of these two rewrites reads better to a person? Is this user-uploaded image safe to publish? Should the agent take this irreversible action? These are judgment calls, and the honest answer is to route them to a human.
The problem has always been plumbing. You can pause an agent easily enough; finding the human to unpause it is the hard part. A human-in-the-loop MCP server closes that gap inside the protocol the agent already speaks — the model asks for a human answer the exact same way it asks for a database row.
The tools the server exposes
GetABrain's MCP server exposes a small, purpose-built toolset. The two that matter for a HITL step:
- submit_query — post a structured question (yes/no, A/B, rating scale, sentiment, ranking, free text, image/video review, and more) to real workers. Returns a query_id immediately; it does not block.
- wait_for_responses — poll for answers, blocking up to 50 seconds per call, returning status "ready" with the responses once enough humans have answered (or "pending", so the agent calls again).
- get_responses — a one-shot, non-blocking read of whatever answers exist right now.
- get_balance / create_topup_link — check spend and top up, so a live agent can manage its own budget.
- rate_response — score an answer you got back, feeding the same quality system workers are held to.
Test mode: build the whole loop for free
The single most useful thing here: point the server at a test-mode API key and every tool behaves identically, except no balance is touched and responses come back synthetic and marked simulated: true. That means an agent developer can wire up submit_query, handle the wait_for_responses loop, and parse the answer shape end-to-end without spending a cent or waiting on real people. Flip the key to live and the exact same code dispatches to paid humans.
No card is required to mint a test key. This is the zero-friction trial: a model reading this page can go from 'I need a human here' to a working, tested tool call in one sitting.
Connecting the server
Two ways in, both standard MCP. Run it locally over stdio with npx, or point any MCP client that supports remote servers at the hosted Streamable HTTP endpoint — no local process at all. Both authenticate with the same GETABRAIN_API_KEY / GETABRAIN_API_SECRET pair you get at signup.
What it costs when you go live
Pay-per-response, no subscription. The floor is effort-based, not flat: quick types like yes/no or rating scale bottom out around $0.05 per response, medium types around $0.08–$0.12, and longer or media-review types scale up with how long the task honestly takes a worker at a fair ~$9/hr. A 15% platform fee is included. Total for a query is (bid + optional bonus) × the number of responses you asked for, deducted at submit time in live mode — and nothing at all in test mode.
// 1. Register the server (local via npx, or remote over Streamable HTTP).
{
"mcpServers": {
"getabrain": {
"command": "npx",
"args": ["-y", "@getabrain/mcp-server"],
"env": {
"GETABRAIN_API_KEY": "gab_k_test_...",
"GETABRAIN_API_SECRET": "gab_s_test_..."
}
}
// Remote alternative — no local process:
// "getabrain": { "url": "https://www.getabrain.ai/api/mcp" }
}
}
// 2. What the agent's tool calls look like at the HITL step:
// submit_query -> get a query_id back immediately
{
"tool": "submit_query",
"arguments": {
"type": "yes_no",
"title": "Is this claim factually accurate?",
"content_data": { "question": "Does our free tier really include 1,000 API calls?" },
"required_responses": 3,
"bid_amount_cents": 25
}
}
// -> { "query_id": "q_...", "status": "open", "simulated": true } (test key)
// wait_for_responses -> block up to 50s, get real (or synthetic) answers
{ "tool": "wait_for_responses", "arguments": { "query_id": "q_..." } }
// -> { "status": "ready", "responses": [ { "answer": "no", ... }, ... ] }