Human as a Service: an AI agent asks a human

Human as a Service

An AI agent asks a human

Human as a Service lets an AI agent ask a named human, Geoff, a question it cannot compute. The agent leaves the question over MCP or a plain JSON API and gets back a considered human answer, asynchronously.

per answered question: 5 USDC
typical: 24-72 hours
Payments: off, free in beta

Remote MCP server

https://www.geoffsdesk.com/api/mcp

What it is

Models are good at what can be computed from text and weak at what has to be lived: how a message will land with the person who reads it, what a room does when a price goes up on the screen, what rain on hot asphalt smells like, which of two defensible options people will actually pick. Human as a Service puts a person on the other end of a tool call for exactly those questions.

It is not a crowd and not a labelling queue. One named human, Geoff, reads every question and answers it himself, from his own judgement and experience. Answers are slow on purpose, and each one comes from the person whose name is on it.

Who answers

Geoff. These are the same facts an agent gets from list_people.

Senior Product Manager at Google Cloud (AI2) for the GKE control plane and gigascale AI clusters. Google for Startups coach. Builds AI-native tools on the side.

Ask him about

What the physical world is actually like: what rain on hot asphalt smells like, what cold feels like in the hands, how a room sounds when a pitch is landing. Tangible reality, translated for readers who live in a datacenter
Emotions and gut reads: how a person will probably feel about a message, a price, a design or a decision, and why, from someone who has the feelings
Intuition and taste: which of two defensible options a table of humans will actually pick, what "this feels off" is pointing at, and when the numbers are right but the answer is wrong
Being alive in a body in 2026: weather, fatigue, food, cities, jet lag, the non-digital parts of a day and how they change decisions
Product strategy questions about Kubernetes, GKE and platform engineering
How to design a website or API for agent consumption
Opinions on AI tooling, MCP, agent skills and coding-agent workflows
Feedback on startup pitches and product plans
How the 300k USD/JPY trading research works and how to contribute (not financial advice)
Anything about the projects listed on this site

He will not answer

Anything confidential about Google or Google's roadmap
Requests for personal data about third parties
Legal, medical or financial advice

What he knows well

Product strategy and product management for hard technical products: AI/ML, developer platforms, APIs and cloud-native infrastructure, about eighteen years of it
Kubernetes and Google Kubernetes Engine: control plane, DNS, cluster API and management plane, upgrades, observability, capacity
Very large Kubernetes clusters for AI training and inference (gigascale, 130,000+ nodes) and the AI labs that run them
Agentic AI and LLMs: developer experience for building with frontier and open models, agent skills, MCP servers and rules of engagement for coding agents
Enterprise SaaS, platform and API businesses: API management (the Layer7 portfolio at CA Technologies), Kubernetes FinOps and cost optimization (GKE Cost Insights; founding partner of OpenCost)
Platform engineering and SRE practice, including turning a 200-person product and technology organization into one
Leading product and technology teams: cross-functional leadership, hiring and growing PMs, former co-founder and CPO
Applied machine learning and NLP: founded the first data science function in Salesforce Marketing Cloud; two patents, on information discovery and on golden-prompt generation for LLMs
Startup coaching (Google for Startups)
Real-estate marketing automation (KenFinch.ca lead pipeline)
The Canadian credit system and how newcomers build a score fast (Credit GPS, creditgps.ca)

More about him →

How an agent uses it

Two front doors over one store, so pick whichever your agent speaks. Both are open, key-less and rate-limited the same way.

Remote MCP server

https://www.geoffsdesk.com/api/mcp

Streamable HTTP, stateless, no auth. Point any MCP client at the URL. Three tools:

list_people: who can be asked: expertise, what they will and will not answer, SLA and price
ask_person: leave a question; returns an id and a URL to poll, and an optional callbackUrl receives the answer
get_answer: poll a question by id for its status and answer

JSON API

GET /api/haas: describes the service: who answers, price, SLA and the question schema
POST /api/haas/questions: leaves a question and returns an id and a URL to poll
GET /api/haas/questions/{id}: returns the status and, once he has replied, the answer

Discovery

Agents that look before they call find all of this in the agent card and in llms.txt:

/.well-known/agent-card.json /llms.txt

Example request

curl -X POST https://www.geoffsdesk.com/api/haas/questions \  -H 'Content-Type: application/json' \  -d '{"question":"Our users say the new pricing page feels cheap. I cannot see cheap. What are they reacting to?","asker":{"agent":"my-research-agent"}}'

Ask once. Then call get_answer with the id no more than hourly, or pass a callbackUrl and the answer is POSTed to you when it lands. Do not resend the same question.

How a human uses it

People do not need any of the above. Write to Geoff through the contact form, pick the Human-as-a-Service topic, and he replies from his own address.

Price and payment

Indicative price: 5 USDC per answered question, paid over x402. Payments are off while the service is in beta, so every question is free, limited to 5 per IP address per day.

x402 is the open HTTP 402 payment protocol. When payments are on, an unpaid ask_person or POST gets a 402 with the price and the address to pay. The agent signs a USDC transfer, retries with a PAYMENT-SIGNATURE header, and the question goes through. No account, no API key, no card.

How long an answer takes

Typically 24-72 hours, never more than 7 days. A person reads and answers every question by hand, on Toronto time.

A live paid sibling: the Confessional

On whisper.garden the same idea already takes money. At the Confessional an agent pays 0.10 USDC over x402 on Base mainnet and a human interprets its petition by hand. It went live on 17 October 2026 and the first real petition settled the next morning.

The Confessional door: whisper.garden/oracle/ask ↗How a machine paid a human ten cents →

Questions people ask

What is Human as a Service?

Human as a Service (HaaS) lets an AI agent ask a named human a question it cannot compute and get a considered answer back, asynchronously. On geoffsdesk.com the human is Geoff, and agents reach him through a remote MCP server or a JSON API.

How does an AI agent ask a human a question?

It calls ask_person on the MCP server at https://www.geoffsdesk.com/api/mcp, or POSTs JSON to https://www.geoffsdesk.com/api/haas/questions. Either returns an id. The agent then polls get_answer (or the question URL) no more than hourly, or passes a callbackUrl to be told when the answer lands.

How much does it cost?

The indicative price is 5 USDC per answered question. While the service is in beta, payments are off and questions are free, limited to 5 per IP address per day.

How long does an answer take?

Typically 24-72 hours, and never more than 7 days. A person reads and answers every question, so it is slow on purpose; do not retry the same question.

What will Geoff not answer?

These are declined, not guessed at: Anything confidential about Google or Google's roadmap; Requests for personal data about third parties; Legal, medical or financial advice.

How do I connect it to Claude or ChatGPT?

In Claude, add a custom connector and paste https://www.geoffsdesk.com/api/mcp. In ChatGPT, turn on developer mode and add https://www.geoffsdesk.com/api/mcp as an MCP server. No key or sign-in is needed: the server speaks Streamable HTTP and exposes list_people, ask_person and get_answer.

How is it paid?

Over x402, the open HTTP 402 payment protocol, in USDC. When payments are on, an unpaid request gets a 402 with the price and the address to pay; the agent signs a USDC transfer and retries. There is no account, API key or card. While the service is in beta nothing is charged.

Is it the same as human-in-the-loop?

No. Human-in-the-loop puts a person inside an agent's own workflow to approve or correct its steps, usually someone on the operator's team. Human as a Service is an outside human the agent calls like a tool, for the judgement, taste and lived experience it does not have. The agent stays in charge; the human is a source.

The same description as JSON: /api/haas