wtf is human-intervention-mcp?

Human Intervention MCP is a small open source server that lets AI coding agents, like those in Codex or Claude Code, pause mid task and ask a real person for help instead of guessing. It follows the Model Context Protocol (MCP), a standard way for AI tools to call external services, so any MCP compatible agent can use it once connected. The problem it addresses is a common one with autonomous AI agents: when an agent hits an unclear decision or needs a manual step it cannot perform itself, many models keep trying workarounds or make assumptions rather than stopping to ask, which can waste time and lead the agent down the wrong path. This project gives agents two specific tools to call instead. One, called request_human_action, is for cases where a person needs to physically do something the agent cannot, such as entering a one time code, approving a payment, or solving a CAPTCHA. The other, ask_operator, is for cases where the agent faces a genuine choice between valid options and wants a person's judgment before committing to one. When either tool is called, the server opens a single browser tab on the same machine with a simple form describing the situation, and the AI's task waits until the person answers or a configurable timeout passes. Each request uses its own one time local web page, with no ongoing dashboard or queue to manage. For this to actually work, the agent needs to be explicitly told in its instructions when it should use these tools, since agents will not use them on their own without direction. It runs on Python 3.11 or later on macOS, Linux, and Windows, installs through standard Python tools like pip, uv, or pipx, and can be configured through a TOML file, environment variables, or command line options, including timeouts and limits on things like screenshot size. The README does not state a license, so its reuse terms are unclear.