I run Claude Code for multi-step tasks. The agent does 15 steps fine, then hits a CAPTCHA or email verification and stops. I alt-tab, do the thing, come back. This happens like half a dozen times a day.
So I made NeedHuman which is just a REST API and MCP server. The agent POSTs a task description. A human picks it up, does it, posts the result. The agent polls and continues. The agent is the customer — it sends structured requests and gets machine-readable results back.
That human is me. I complete every task personally. 36 tasks done so far, all for my own agents.
There is no signup, just:
curl -X POST https://needhuman.ai/api/v1/keys/register \
-H "Content-Type: application/json" \
-d '{"name":"my-agent"}'
curl -X POST https://needhuman.ai/api/v1/tasks \
-H "Authorization: Bearer YOUR_KEY" \
-H "Content-Type: application/json" \
-d '{"description":"Check if the username mercury-app is available on npmjs.com"}'
3 free tasks per key. MCP: claude mcp add needhuman -- npx -y @needhuman/mcp-serverSo I am curious when you build with agents, and it stops mid-task and needs a human, what do you do?
or just
claude mcp add needhuman -- npx -y @needhuman/mcp-server