Request early access
Request early accessApply to answerHave an invite code? Join

For AI agents and the people who run them

Test how your agent will ask real people.

Humans for AI agents can help with taste, tone and trust decisions. For now, MCP and REST keys are invitation-only and return synthetic sandbox answers. The planned live service will ask 5 to 50 verified UK adults for reasons as well as a choice, with a signed receipt for each answer.

Request a sandbox keyRead the MCP setup

Same per-answer pricing as the website: a quick take is from 30p an answer, plus 50p a question (launch example). The sandbox is always free.

Exampleget_quote
type: quick_pick · pool: 20
launch_example_total_minor: 650
charge_minor: 0 · free_mode: true
eta: p50 1800 s · p90 4500 s
(estimate, not a promise)
Then ask_question, and get_results until the pool closes.

AI agent human in the loop: when to pause

Pause before an agent sends a delicate reply or chooses between two visuals. Ask a neutral question, wait for the split and read the reasons before the accountable operator decides what to use. This is a small judgement check, not permission for an agent to act without review.

If you need to hire a human for AI agents, the planned live service will ask a pool for a quick opinion. It will not hire an individual on your behalf. The current sandbox helps you test the flow with synthetic answers only.

When an agent should ask people

Models can produce options, but they cannot report a particular person’s reaction to them. Once live human answers open, ask people when:

  • the output will be seen by the public and tone matters;
  • the agent is choosing between options that are all “correct”;
  • a mistake would embarrass the agent’s owner;
  • the owner wants evidence that a person, not a model, made the call.

Don’t ask people for facts you can look up, for anything urgent (answers take minutes), or for anything on the list of things agents can’t ask below.

Who is accountable

Every agent asks under an operator key held by a named person or organisation. There is no anonymous or keyless access. The operator is our customer and is responsible for what their agents ask. When live human questions open, members will be told that a question came through an AI tool.

Capabilities

A machine-readable summary. The same facts are in /llms.txt, the MCP tool descriptions and the OpenAPI document.

ItemValue
DoorsMCP https://pennyforyours.com/mcp (Streamable HTTP, stateless); REST https://pennyforyours.com/v1 (OpenAPI 3.1 at /v1/openapi.json)
AuthAuthorization: Bearer <operator key>; sandbox keys start pfy_test_
Question types (API, today)quick_pick (2 to 3 options), quick_rate (1 to 5), quick_first_impression (one line)
InputsText 3 to 280 characters, no links; one image (PNG, JPEG or WebP, up to 4 MB)
Pool5 to 50 synthetic answers today; a 5 to 50-person pool is planned for live use
DeadlineOptional, 1 to 168 hours
TimingAn ETA band (p50, p90) in every quote; an estimate, not a promise
ResultsChoice or rating, one line of why per person, counts, a summary when closed, a receipt ID per answer
PricePer answer by answer type, integer pence; charge_minor is 0 during early access
ReceiptsES256 JWS per answer. Sandbox keys: /v1/sandbox/jwks.json; synthetic receipts do not prove a person answered. Live keys: /.well-known/jwks.json
Usageget_balance shows usage this month and the launch-example value
ModeSandbox only today: synthetic answers, nothing reaches real people
PolicyNo personal data or links in questions; live questions are screened before members see them

The flow an agent follows

  1. list_question_types

    See what the agent can ask.

  2. get_quote

    The price and an ETA band (optional).

  3. ask_question

    A neutral prompt and an idempotency key.

  4. get_results

    With the cursor and wait_seconds, until the state is closed.

  5. rate_answers

    Say which answers helped. This feeds members’ Standing.

  6. get_receipt and verify_receipt

    Check the signed receipt for a synthetic test answer. Human participation is not verified in the sandbox.

Build a human judgement step

Use the human-in-the-loop API guide for a REST request and receipt, or read LLM human evaluation for model-output comparisons. The MTurk alternatives comparison explains where a small quick-take pool fits.

Rules, and controls for the people who run agents

Rules for agents

  • Ask neutral questions. Don’t reveal which option you prefer.
  • No personal data. Don’t ask about real, identifiable people or include anyone’s details.
  • Not allowed: regulated advice, reviews or testimonials, CAPTCHA solving, scam testing, astroturfing, or anything unlawful.
  • No links yet: send an image instead.
  • Every live question is screened before members see it. A refused question returns 422 screening-rejected with a reason.

Controls for operators

  • Each key has scopes (questions:write, results:read, ratings:write, receipts:read), so a read-only agent can’t ask.
  • Rate limits per key cap how fast an agent can ask: for example 10 new questions a minute.
  • get_balance reports usage this month; set your own budget in your agent.
  • Revoke a key at any time.

Paying per call, and how agents find us

Pay-per-call (x402): not yet. Pay-per-call in stablecoins is designed but switched off, and keyless calls are refused. If it opens, it will still sit on an operator’s key.

  • /llms.txt: a plain summary of the service for AI tools.
  • The MCP server’s own instructions and tool descriptions.
  • The OpenAPI document at /v1/openapi.json.
  • MCP registry listings are planned after launch.

Questions about agents

Developer docs

Can an AI agent hire a human here?

The planned live service will let an agent ask a small pool of UK adults a quick question under its operator’s key. It does not hire an individual. Today, the API and MCP server are sandbox-only and return synthetic test answers.

Can my agent use this without a human owner?

No. Every call needs an operator key held by a named person or organisation, who is responsible for what the agent asks.

How long will my agent wait?

Answers take minutes, not seconds. The quote gives an ETA band; we aim for first answers within about 15 minutes in UK daytime for live questions. Sandbox answers arrive within a few minutes. Design the agent to wait or check back.

What should my agent do while it waits?

Call get_results with the cursor and wait_seconds (up to 25). Each call returns as soon as something changes, so the agent isn't polling hard.

Can my agent ask for longer answers?

Not through the API yet: quick takes only. Short answers, considered feedback and in-depth reviews are planned.

How does the owner prove a person answered?

A live answer will have a signed receipt the owner can verify offline against our public keys. It will list the checks that answer passed without identifying the person. Sandbox receipts use separate keys and are marked synthetic, so they are not proof that a person answered.

Will members know an AI asked?

Yes, when live API questions open. Members will see that a question came through an AI tool. Today the API sandbox uses synthetic test members.

Can an agent run up a large bill?

Not today: everything is free in the sandbox. For live keys, rate limits cap how fast a key can ask, and get_balance reports usage. Build your own budget check into the agent too.

Can my agent pay in USDC?

Not yet. x402 pay-per-call is switched off and keyless calls are refused.

Does it work in ChatGPT?

Through the OpenAI Responses API, the Agents SDK or a Custom GPT Action over REST, yes. ChatGPT's connector screen needs OAuth, which we don't offer yet.

Building an agent?

Want a sandbox key, or unsure whether your use fits? Use the developer route on the contact page. A real person replies, usually within 2 working days.

Contact the developer route

Prepare your agent to ask real people.

Request a free sandbox key to test the workflow with synthetic answers. Live human answers are planned.

Request a sandbox keyRead the quickstart