The full upstream README, mirrored here for reference. Install config, tool schemas, adoption signals, and an original overview live on the Hunch listing page.
Hunch gives your AI assistant calibrated judgments about text. For each text you send, it returns the probability that the answer to a yes/no question is yes, one option from your own list, a score on your own scale, or several yes/no answers at once. The answers come from Jev, TypeSafe's System One model, which returns numbers instead of generated text, so your assistant can sort, filter and count them.
The server is hosted at https://hunchsheet.app/mcp (Streamable HTTP). This repo holds the install files for Claude Code and Gemini CLI. Setup for Claude, ChatGPT, Cursor and VS Code is at hunchsheet.app/connectors.
| Tool | Returns |
|---|---|
hunch_ask | The probability of yes, 0 to 1, for each text |
hunch_pick | One option from your list for each text, with its confidence |
hunch_score | A position on your levels (0 to n-1), the most likely level, and its confidence |
hunch_multi | One probability per question for each text, up to 10 questions |
hunch_balance | Credits left on your key |
Good for leads, support tickets, reviews, survey answers, emails and any list of short texts you want judged the same way. It judges text only: no arithmetic, counting or dates. English works best. Put the definition inside the question, for example "Can this person approve a software purchase without asking someone else?" instead of "Is this a decision maker?".
The form at hunchsheet.app emails a key with 100 free rows. After that, 5,000 rows cost $29 and credits never expire. One credit per answered text per question. Blank texts and repeats in the same call are free.
Set your key before starting Claude Code:
The plugin adds the server and a skill that tells Claude when to use each tool. Without the plugin:
Gemini CLI asks for your key during install and stores it as a sensitive setting.
https://hunchsheet.app/mcp, then sign in with your key when Claude opens Hunch's page.The texts and question in a tool call go to the Hunch server and to TypeSafe to compute the answer, and are not stored. Details at hunchsheet.app/privacy.
The files in this repo are MIT licensed. The Hunch service has its own terms.