For developers and AI builders
Add a human check to your app or AI workflow.
When your product needs a person to judge something (which output feels better, whether a reply would calm a customer down, whether a generated image looks real), it can ask verified people in the UK and get structured answers back, each with a signed receipt.
Free sandbox. Later from 30p an answer, plus 50p a question (launch example).
Why builders add people
Models grade models well enough for many things, but not for how something feels to a person seeing it for the first time. Labelling services sell volume; we sell a small number of verified, reasoned human judgements, each with proof of where it came from.
Researchers have found that many paid crowd workers already use AI to write their answers (one study). Verified people with signed receipts are the gap PennyForYours is built for.
Questions builders ask
Questions where people's own reactions beat a guess. Anything marked Later is not available yet.
“Which of these two replies sounds more natural?”
from 30p an answer
“Our agent wants to send this email to a supplier. Would you send it as it is: yes, no, or yes with a change?”
from 30p an answer
“Our bot drafted this reply to a complaint. Would it calm you down or wind you up?”
from 30p an answer
“Would this push notification make you open the app, or turn notifications off?”
from 30p an answer
“How warm does this AI-written birthday message feel, 1 to 5?”
from 30p an answer
“Is this generated product photo realistic? Tap anything that looks wrong.”
from 30p an answerLater
“Write the question you'd type into a support bot about this bill.”
from 80p an answerLater
Launch example Launch examples, per answer, plus 50p a question. Early access is free.
What you get back
A made-up example at launch prices, to show the shape of an answer: the split, the reasons behind it and a short summary.
Which of these two replies sounds more natural?
- “The first sounds like a person who cares.”
- “The second reads like an automatic email.”
- “Short and clear. I would trust the first.”
What you can build
- Pre-publish checks. Before your app sends or publishes AI-generated copy, ask five people whether it sounds natural or would offend.
- Model and prompt comparisons. Blind pairwise preference between two outputs, with a line of why from each person.
- Image checks. "Which image feels more like a cosy autumn café?"
- Escalation inside an agent. When an agent is unsure, it asks people instead of guessing. See AI agents.
How it connects
| Door | Details |
|---|---|
| MCP server | https://pennyforyours.com/mcp (Streamable HTTP, stateless, bearer key). Works with Claude Code, Claude Desktop and claude.ai, Cursor, VS Code, OpenAI Codex and other MCP clients. Setup per client |
| REST API | https://pennyforyours.com/v1 with an OpenAPI 3.1 document. Long-polling or a resumable event stream for results. API reference |
| Receipts | A signed JWS (ES256) per answer, verifiable offline against our public keys. How receipts work |
| ChatGPT and OpenAI | Through a Custom GPT Action (REST with a bearer key), the Responses API MCP tool or the Agents SDK. ChatGPT's own connector screen needs OAuth, which we don't offer yet. Details |
Status, plainly
- Sandbox only. Keys start
pfy_test_, synthetic test members answer within a few minutes, and nothing reaches real people or costs anything. - Keys by invitation during early access.
- Quick takes only through the API today.
- Not yet: live keys, webhooks, OAuth for MCP clients, keyless pay-per-call.
Rules for builders
Every call needs an operator key with an accountable owner; there's no anonymous access. Not for regulated advice, collecting personal data, writing reviews, solving CAPTCHAs or testing scams. Members are told when a question came through an AI tool.
What it costs
Launch exampleA quick take is from 30p an answer, plus 50p a question for screening and the summary.
- 20 quick takes ≈ £6.50
The same per-answer pricing as the website. The sandbox is always free, and quotes show a charge of 0 during early access. Pricing for developers
How it works
Ask
Write your question, add an image if you need one, and choose how many people answer, from 5 to 50.
Verified people answer blind
They see your question without your name, and nobody sees anyone else’s answer first.
Get a summary and the reasons
See the split, a line of why from each person and a short summary.
Use cases for developers and ai builders
Each page shows the questions developers and ai builders ask, what you get back and the honest limits.
Questions from builders
How do I get a key?
Keys are issued by invitation during early access. Use the developer route on the contact page and tell us briefly what you're building. Sandbox keys start pfy_test_ and are shown once.
Do sandbox answers come from real people?
No. Synthetic test members answer sandbox questions, every answer is marked synthetic, and sandbox receipts never verify as proof of a person. The sandbox shows you how the API behaves, not what people think.
Can I use it from ChatGPT?
Not through ChatGPT's own connector screen yet, because it needs OAuth and we use API keys. You can use a Custom GPT Action over our REST API with a bearer key, or the OpenAI Responses API and Agents SDK, which can send the key as a header.
Does it work with Claude?
Yes. Add it in Claude Code with one command, or as a custom connector in Claude Desktop and claude.ai with your key as a request header.
How fast are results?
Sandbox answers arrive within a few minutes. For live questions we aim for first answers within about 15 minutes in UK daytime, and the quote returns an estimate. Design your flow to wait, using long-polling or the event stream.
What can a receipt prove?
That an answer with a specific hash was given to a specific question and image through an assignment we issued, at a stated time, with the checks listed in the receipt. It doesn't identify the person, and browser-reported signals are supporting evidence, not proof.
Are there rate limits?
Yes, per key and shared between REST and MCP. For example, 10 new questions a minute and 120 result reads a minute. The sandbox page lists every limit.
Can I use the answers to train a model?
Not yet settled. Today answers are for your own decisions and evaluation. Tell us if training use matters to you; licence terms for that would need to be agreed and members told.
What will it cost when it goes live?
Per answer, by answer type: a quick take is from 30p an answer, plus 50p a question (launch example). Quotes return integer pence and show a charge of 0 during early access.
Want a sandbox key?
Tell us briefly what you're building, or whether your use case fits, through the developer route on the contact page. We'd rather hear about it before you build. A real person replies, usually within 2 working days.
Add a human check.
Free during early access. No card needed.
Free during early access. We review requests by hand. Have an invite code? Join