A model will generate ten variants and then guess which one a person prefers. Handing that judgement to a panel and picking the result back up in the same conversation turns it from a guess into data — A/B preference, ratings, rankings, or a human gate before an agent ships something irreversible.
A server over Datapoint AI, which recruits human annotators. From the conversation you design a survey from a plain-language description, launch it to a panel, and read status and aggregated results as they come in. Text, images, audio and video are all supported as material.
- A survey designed from a natural-language description, before anything is launched
- The survey launched to a panel of real people
- Local images, audio or video uploaded so they can be used as survey material
- Status, progress and aggregated results while it runs
- Raw per-annotator responses, paginated, when the aggregate is not enough
- All your surveys listed
- A running survey paused and resumed — in-flight responses keep arriving while it is paused
- A survey cancelled permanently, with unused reserved credits refunded
- Account balance checked, and a checkout link to top it up
Uv on your PATH, and a Datapoint AI account — the first run opens a browser to authenticate. Surveys are paid for in credits, and cancelling refunds the reserved credits you did not use.
One command — uvx --from git+https://github.com/impel-intelligence/datapoint-mcp.git datapoint-mcp