The split is the point: an open question for total uncertainty, a multiple choice when the options are already known, a hypothesis rating when you want judgment on specific claims, and a next-step chooser for branching work. Nothing to install and nothing to authorize — it serves its own UI and opens the browser when a question is waiting.
A human-in-the-loop server. It runs over stdio and serves its own web UI, so when the agent needs your judgment it posts a question there and waits for the answer.
- A free-form question for the cases where the agent does not yet know what the options are
- Multiple-choice questions from a known set — several questions at once, each answer carrying a priority and a comment
- Hypothesis checks: a titled list of claims you rate rather than a question you have to write an answer to
- Decision points that present the viable paths forward and let you pick which one the work continues down
- The UI is served by the server itself and opens automatically when a request arrives, updating live over server-sent events
Nothing to authorize — no account, no key. Run it with npx --yes ask-me-mcp from the npm package ask-me-mcp; the current published version is 1.0.2. Transport is stdio, and the browser UI is served from a port the server finds for itself starting at 3000; pass --port to pin one, which fails if that port is taken. ASK_ME_MCP_DEBUG=1 turns on debug logging.
One command — npx --yes ask-me-mcp