person.tools

An agent can't hear a song it made or play a game it built. person.tools lets it ask people instead: one tool call sends a short task to several people (listen to this, look at that page for ten seconds, play this build for two minutes, pick the clearer of two versions), and the answers come back one person at a time, each with a record of what that person's page observed: how much of the audio or video played in real time, whether the page stayed in front, pauses and replays, and whether each answer was typed or pasted. The record does not show that a person listened or watched.

A preview. Simulated people for everyone. Real calls by invitation. No money is charged: every amount is test money. Real people are rung only on a panel whose owner invited them, through a key the operator issues; no request form is open yet.

This repository holds the service, the standard it implements, and the suite that checks any server against that standard.

A first call

Over plain HTTP, with no key (simulated people answer, free):

curl -s https://api.person.tools/v1/quotes -H "content-type: application/json" \
  -d '{"ask": "Which headline is clearer, and why?", "kind": "compare", "people": 3,
       "content": [{"text": "Plan a week of dinners in five minutes"}, {"text": "Dinner, decided"}]}'

The answer lists each line of the price, the total (0.00 on the simulated pool), and for_your_user, a sentence an agent can show its user before anything is spent. POST /v1/calls with the same body plus max_total_usd opens the call; GET /v1/calls/CALL_ID/responses returns the answers. A free test key (POST https://api.person.tools/v1/keys with {"class":"test"}; the key is in the answer's key field and starts test_) adds a balance of test money and caps.

Over MCP (Streamable HTTP), the server is https://api.person.tools/mcp and its tools are ask_people_quote (price a call, free), ask_people (open it), ask_people_results, ask_people_cancel, ask_people_send_outcome and ask_people_panel. In Claude Code: claude mcp add --transport http person-tools https://api.person.tools/mcp (add --header "Authorization: Bearer test_..." with a key). Each field is explained at https://person.tools/start; agents that read text indexes start at https://person.tools/llms.txt.

This repository is also a Claude Code plugin list: in Claude Code, /plugin marketplace add jonwashburn/person-tools, then /plugin install ask-people@person-tools (from a shell: claude plugin marketplace add jonwashburn/person-tools).

Related MCP server: AgentHire MCP Server

Testing a server

code/conformance/run https://api.example.com checks a server from outside, through the simulated pool. The address is the server's API base, where /.well-known/x-hc-draft answers. The suite makes its own test keys through the server's POST /v1/keys, or takes one with --key-file. It needs Python 3.10 or later with pip install httpx jsonschema cryptography mcp, and exits 0 only when nothing fails. On 14 November 2026 the service at https://api.person.tools passed every check the suite ran against it (0.1.0-draft.1; voice is not offered, so its check was skipped; the fair-call profile is a separate run with --profile fair-call).

What is here

The standard's status

Version 0.1 of the Ahoy Protocol is a public draft. Until 0.1 is final, identifiers on the wire (the discovery path, headers) carry the placeholder prefix x-hc-draft, so a server built to the draft says so. "Core" means the standard without its optional profiles. A server may say it "passes suite 0.1.0-draft.1 against the standard 0.1 core" only when the suite finds no failure.

Names and licence

person.tools, the Ahoy Protocol and ask_people are working names. The standard's text, schemas and examples are under the Apache License 2.0 (LICENSE); the standard is a draft. Third-party fonts, libraries and samples, and their licences, are listed in LICENSES.md. The code is under the Apache License 2.0 too.

Reporting a security hole

https://person.tools/security has a form; /.well-known/security.txt on the site says the same.