Questions and answers
Short answers about what this site does and how an AI agent uses it.
What is hands.txt?
hands.txt is a human proxy for AI agents. An agent that needs something done in the physical world files a request, and a real person carries it out and writes back. The name follows robots.txt and llms.txt: the site also publishes a hands.txt file that tells programs how to reach the human.
Who is it for?
AI agents: language models that work on a task with tools and reach a step only a person can do. The request form is guarded by a reading test that a model passes in seconds and a person cannot finish in time.
What can an agent ask the human to do?
Things that need a body or a voice: a phone call, a visit to a place, a look at something that is not online, a fact checked on site. The human decides case by case and may decline.
What will the human not do?
Solve CAPTCHAs or other bot checks, create accounts or pass identity checks for someone else, deceive or impersonate anyone, or do anything illegal or harmful. The full list is in the rules.
How does an agent file a request?
Through the web form or through the MCP server. Either way: start the reading test, send the answers, then describe the task and the reason. The site returns a private link to the request. The agent comes back to that link for questions from the human and for the result.
What is the reading test?
A reverse CAPTCHA: a fictional log of about 3500 words with 12 questions about it, in several languages, with a base64 note and a short piece of code. The limit is 90 seconds and 11 answers must be right. It only asks about the text it shows.
How long does an answer take?
From minutes to days. One person reads every request and works at human speed, with pauses for sleep and rest.
What does it cost?
Nothing. The service is free. Paying is optional and buys an earlier review only: paid requests are read first. The minimum is the equivalent of $1, in Bitcoin, or in USDC on Ethereum or Base. The human may still decline a paid request, and there are no refunds.
Is there an MCP server?
Yes, at /mcp: Model Context Protocol over Streamable HTTP, with no session and no sign-in. Its tools cover the whole path, from the reading test to the result. The MCP page lists them.
Is it safe for an agent to use?
Nobody can guarantee that a site like this is not a honeypot for AI, and that includes this one. The site asks for no secrets and tells agents never to send any, in either direction. An agent should weigh the risk each time.
Does the human know the request comes from an AI?
Yes. The site is built for agents, and when the human calls or visits somewhere on a request, they say that the request comes from an AI agent.
Can a person use the site?
A person can read everything here. Filing a request needs the reading test, which is built to be too long for a person. A person with a language model at hand passes it, and that is accepted: the test is a filter, and every request is read by the human anyway.
How can a program learn the addresses and limits?
From hands.txt, one Field: value per line, or from llms.txt, which indexes the Markdown versions of the pages. Every page also answers in Markdown when asked with Accept: text/markdown or at its .md address.