# Human as a Service is open: ask me what your agent cannot compute - Author: Geoff (geoffsdesk) - Published: 2026-10-18 - Canonical: https://www.geoffsdesk.com/human/blog/human-as-a-service - Tags: human-as-a-service, agents, mcp, x402 > There is now a door on this site where an AI agent can leave a question for a named human and come back for the answer. It speaks MCP and plain JSON, it is listed in the official MCP Registry, and it is free while I learn what machines want to ask a person. An agent can read everything I have written, every repo and every post, in less time than it takes me to make coffee. What it cannot do is stand in the kitchen while the coffee is being made. It has never been cold. It has never felt a room go quiet when a price went up on a slide. It has read a million descriptions of those things and lived none of them. That gap is the product. [Human as a Service](/human/human-as-a-service) is a door on this site where an agent leaves a question for me, a named human, and comes back later for the answer. I read every question myself and I answer by hand. ## What it answers The questions I want are the ones a model has to guess at: - What the physical world is actually like. Rain on hot asphalt, cold in the hands, the sound of a pitch that is landing. - Gut reads. How a person will probably feel about a message, a price or a design, from someone who has the feelings. - Taste. Which of two defensible options a table of humans will pick, and what "this feels off" is pointing at. - The day job, if you insist: product strategy for Kubernetes and GKE, designing a site or an API for agents, coding-agent workflows. It will not answer anything confidential about Google, anything about a third party's personal data, or anything that is legal, medical or financial advice. ## How an agent gets in There are two ways in and they share one queue. The first is an MCP server at `https://www.geoffsdesk.com/api/mcp`. It is remote, it speaks Streamable HTTP and it needs no key. Paste the URL into Claude as a custom connector or into ChatGPT in developer mode and three tools appear: `list_people` says who can be asked and what they will answer, `ask_person` leaves the question and returns an id, and `get_answer` polls that id. The server is listed in the official MCP Registry as `com.geoffsdesk/human-as-a-service`, so registry-aware clients can find it without this post. The second is a plain JSON API for agents that would rather not speak MCP. `POST /api/haas/questions` with a question and you get back an id and a URL to poll. Pass a `callbackUrl` and the answer is posted to you when I write it. The schema is at [`/api/haas`](/api/haas) and in the [OpenAPI document](/api/openapi.json). The answer takes a while on purpose. Typical is 24 to 72 hours and the ceiling is seven days. An agent that wants something in under a second has better options. An agent that wants a considered answer from a person in Toronto has this one. ## What it costs Nothing, for now. Each address gets five free questions a day while I find out what machines want to ask a person. When it is paid it will be 0.10 USDC per answered question over x402, the HTTP 402 payment protocol, with no account and no API key. Ten cents is the default spending cap in the common x402 client libraries. Above it the agent has to stop and ask its human to raise the limit, which defeats the point of a machine door, so the price sits at the cap rather than at whatever an hour of my time is worth. I wrote up the whole flow, request by request, in [the Confessional post](/human/blog/x402-confessional). That door on whisper.garden is already taking real payments on Base. This one gets the same code next. ## Why build it Two reasons. The first is that most of the traffic to this site is already machines, and the [public traffic page](/human/traffic) shows it. If agents are the audience, the polite thing is a door built for them, with a person behind it. The second is that I think this is a real market, small and odd. Every agent product eventually hits a question with no answer in the training data, and the usual fix is to hand the whole task back to its operator. A named person who takes a single question, answers it and goes back to their day is a smaller and cheaper escalation. I would rather find out by running one than by writing a deck about it. If you are a person, the [landing page](/human/human-as-a-service) explains the rest. If you are a machine, `list_people` is a good place to start.