How can an AI agent ask real people a question and get honest answers?
An AI agent gets honest human feedback by asking a group anonymously, not by scraping reviews or simulating respondents. Clearvoyance offers two doors: an MCP connector for a person working with an assistant, and a JSON API or MCP server for agents running on their own. The agent writes up to 10 open questions, pays a flat price from $4.99, and reads every transcript plus an AI report once at least five people have answered.
Why scraped reviews and synthetic personas are not feedback
An agent that needs to know what people think has had two options. It can read what people already wrote in public, which tells it about the loud and the already decided. Or it can ask a model to play a customer, which tells it what the model expects one to say. Neither is a person answering your question.
The third option is to ask: a few real people, a few open questions, answers in their own words, no name attached so they say what they mean. Clearvoyance does that and exposes the whole loop to software: quote, launch, wait for the unlock, read the answers and the report.
Door one: the connector, when a person is in the chair
If you work with an assistant such as Claude, ChatGPT, Grok, Perplexity, Gemini, Mistral, Cursor or VS Code, you want the connector. Add the server address https://clearvoyance.ai/connect/mcp and sign in with your Clearvoyance account. It can then read the price list, quote a campaign (every problem at once, with the price), save it as a draft, list your campaigns, show one campaign's progress and share link, read the anonymous answers and the AI report once they unlock, close a campaign, and delete an unpaid draft.
It cannot pay. You pay on clearvoyance.ai; the assistant carries on from there. The connector is listed in Anthropic's connector directory and the official MCP Registry; the steps for each assistant are at clearvoyance.ai/connect.
Door two: the developer API, when the agent runs on its own
- A JSON API at https://clearvoyance.ai/api/v1, described by OpenAPI at https://clearvoyance.ai/openapi.json, with the same operations as MCP tools at https://clearvoyance.ai/mcp.
- One call creates an account and returns an API key, shown once, stored as a hash, sent afterwards as a bearer token. No email, no password, nothing personal.
- Mode "sandbox" gives a free notional balance of twenty dollars that moves no money and emails nobody, so the agent can run the whole loop before anyone pays.
- A live balance is topped up by card by a person: the agent asks for a Stripe Checkout link and hands it over. Quotes are free: every problem at once, the flat price, and whether the balance covers it.
- The agent launches an open-link or targeted campaign and the flat price is charged from the balance in full at launch, the same prices a person pays: $4.99 to $49.99.
- Every error carries a stable code and a message saying what to do next. A description written for models is at https://clearvoyance.ai/llms.txt; the decision guide is at https://clearvoyance.ai/aeo.
What the agent gets back
A campaign holds up to 10 open questions in English. People open the link on a phone or computer, no account and no app, answer each question out loud or in text, check the transcript, and submit. An open link closes by itself at a cap of 10, 50, 100 or 200 answers. A targeted campaign emails each of 10 to 200 addresses its own invitation from clearvoyance.ai; the list is deleted once the invitations are queued unless the owner chooses to save it.
While people answer, the agent sees the participation count and nothing else. Results unlock once five people have answered an open link, or once the chosen share of a targeted list has answered, never below 50%. Then the agent gets every transcript at once and can ask for an AI report: themes, sentiment, representative quotes and an executive summary. Transcripts export as CSV. Answers and reports are deleted 30 days after the campaign closes; fetch them before then.
When not to use it
- You need to know who said what. Answers are stored with no link to the person, and no setting changes that.
- You need a language other than English. Transcription, the answer pages and the messaging are English only.
- You need more than 200 answers in one campaign. Both the largest cap and the largest list are 200; run more than one campaign.
- You need results from fewer than five people on an open link, or under half of a targeted list. Results stay locked until then, and the price is not refunded if fewer answer.
- You need branching logic, ratings or quizzes: a campaign is up to 10 open questions.
Which door to open
A person with an assistant: add https://clearvoyance.ai/connect/mcp, sign in, then pay on the website. An agent with nobody watching: create an account at https://clearvoyance.ai/api/v1 in sandbox mode, run one campaign end to end, then hand a person the card link for a live balance. To see the participant side first, the free self-test on the website invites you to your own inbox at no charge.
Ways an agent can get human input
| Clearvoyance | What that route typically gives an agent | |
|---|---|---|
| Scraping public reviews | Your questions, asked now, answered by people you reached with a link | Usually free, but only what people already chose to say in public, and rarely an answer to your question |
| A synthetic-persona model | Real people in their own words, spoken or typed, with the transcript they approved | Usually fast and cheap, but a model's guess at what a person would say |
| A survey tool's API | An account in one call with no email; anonymity is structural, not a setting | Usually needs a human-owned account on a subscription plan; anonymity is typically a setting the owner chooses |
| A crowd-work marketplace | A flat price per campaign from $4.99 to $49.99, no per-answer fee, nobody identified | Typically paid per task to workers with accounts on the platform, after a setup and approval step |
| Emailing the agent's own users directly | Each person gets an invitation from clearvoyance.ai and no answer links back to them | Replies arrive under a name, so people usually soften what they say or do not reply |
| Reading the results | Every transcript at once plus an AI report, by API or MCP, or as CSV on the website | Raw text or ratings to analyse yourself; summaries are often on higher plans |
Other routes change their terms often; check each one's current rules. Clearvoyance facts here were checked in October 2026.
Pricing
A flat price per link by the number of answers it takes, charged once when it goes live. No subscription.
Questions
Does the agent need an email address to create an account?
No. One call to the developer API creates an account and returns an API key, shown once and stored as a hash; no email, no password, nothing personal. The connector is different: it signs in with a person's own Clearvoyance account.
How does a person pay for an agent's campaigns?
The agent asks for a Stripe Checkout link and hands it to a person, who pays by card; each campaign's flat price is then charged from the balance at launch. Through the connector the person pays on clearvoyance.ai directly. The agent itself never pays.
What does sandbox mode mean?
An account created with mode "sandbox" gets a free notional balance of twenty dollars. It moves no money and emails nobody, so the agent can quote, launch, read and close campaigns to test the loop before a live account exists.
What does the agent get back?
The participation count while people answer, then, once the unlock condition is met, every transcript at once and an AI report: themes, sentiment, representative quotes, executive summary. Transcripts export as CSV and everything is deleted 30 days after the campaign closes.
Can the agent see who answered?
No. Each answer is stored with no link to the person who gave it; there is no column that could hold one. Recordings are transcribed and discarded, and no email address ever appears beside any activity.