Hunch for AI agents
Hunch turns a batch of short texts into typed numbers: a probability, a pick from your options, or a score on your scale. An agent calls it instead of reading hundreds of texts itself. One call judges up to 500 texts. There is a keyless sample, a free 100-row key with no email, and a checkout link that a person pays.
When to call Hunch, and when not to
Call Hunch when:
- More than about 25 texts need the same judgment: triage, lead scoring, tagging, routing, filtering, survey coding.
- You need a comparable number per text, not a paragraph per text.
- The job repeats and the answers should stay consistent from run to run.
- You want the texts out of your context window.
Do not call it when:
- There are fewer than about 25 texts. Read them yourself.
- The task needs arithmetic, counting, date comparison or facts that are not in the text. Hunch reads the text and nothing else.
- The text is not English and accuracy matters. English works best.
- The text holds card numbers, health records, government ID numbers, passwords or API keys.
The tools
| Tool | What it returns | Costs |
|---|---|---|
hunch_ask | A yes/no probability, 0 to 1, per text | 1 credit per text |
hunch_pick | One option from your list per text, with confidence | 1 credit per text |
hunch_score | A position on your ordered scale (2 to 10 levels) per text | 1 credit per text |
hunch_multi | Up to 10 yes/no probabilities per text in one call | 1 credit per question per text |
hunch_quote | Credits and dollars a job needs, before you run it | Free, no key |
hunch_balance | Credits left on the key | Free |
hunch_get_key | A key with 100 free rows, in the result, no email | Free, keyless connection only |
hunch_buy_credits | A checkout link for the person to open | Free until the person pays |
Blank texts and texts repeated inside one call are free. Put the full definition of yes, or of each option or level, inside the question: the model sees the text and the question, nothing else.
Try it with no account
Connect https://hunchsheet.app/mcp/try (install lines for every client are on the install page). Call any judging tool with up to ten texts. The free sample is ten texts a day per network address.
curl -s https://hunchsheet.app/mcp/try -H 'content-type: application/json' -d '{"jsonrpc":"2.0","id":1,"method":"tools/call","params":{"name":"hunch_ask","arguments":{"texts":["Please refund me now, this is the third time","Thanks, all good"],"question":"Is the writer asking for a refund?"}}}'The result carries a table for you to read and the same data as structuredContent.
Get 100 free rows with no email
Agents have no inbox, so the key comes back in the response. Over MCP, call hunch_get_key and pass the key as api_key to the other tools. Over HTTP:
curl -s -X POST https://hunchsheet.app/trial -H 'content-type: application/json' -d '{"source":"agent"}'Each network address can take two keys a day. Treat the key like a password: it spends credits.
How an agent buys credits
A person pays. An agent never holds card details and cannot pay for anyone. What an agent does is make the purchase one click:
- Price the job.
hunch_quoteorGET /v1/quote?rows=Nreturns the credits needed and the dollars, with the free rows and the plans applied. - Get the link.
hunch_buy_creditsorPOST /v1/checkout(with the key) returns a Stripe checkout link that adds credits to that same key. The link carries a one-hour ticket, not the key. - Hand it over. Show the link to the person. Credits land on the key the moment payment clears.
| Plan | Price | Credits | Per 1,000 |
|---|---|---|---|
| Free key | $0 | 100 | $0 |
| Starter | $29 | 5,000 | $5.80 |
| Top-up (same key, after Starter) | $19 | 25,000 | $0.76 |
Credits never expire and there is no subscription. The same numbers are in pricing.json.
Paying by card on Stripe's page is the only payment method today. There is no x402 or agent-wallet payment yet.
When the key runs out
The judging tools say which texts were skipped and include the checkout link in the result. Over HTTP, POST /v1/ask with the header X-Hunch-Agent: your-agent-name answers 402 when nothing could be answered for lack of credits. The body's payment field holds the offer, and the same link is in the X-Hunch-Checkout-Url header. Without the header the response stays a 200 with per-text errors, which is what Google Sheets and Excel expect.
HTTP API
Everything above is also plain HTTP, described in openapi.json, and listed at /.well-known/api-catalog.
curl -s https://hunchsheet.app/v1/ask -H 'Authorization: Bearer hunch_your_key' -H 'content-type: application/json' -d '{"kind":"score","question":"How does the reviewer feel about the product?","options":["angry","disappointed","neutral","happy","delighted"],"rows":["Support fixed it in five minutes, impressed"]}'Data
The texts you send go to Hunch and to its model provider, TypeSafe, to compute the answers, and Hunch does not store them. Details are on the privacy page.
Questions
Can an agent buy credits without a person?
Which agents does it work with?
How is this different from asking the model I am already running?
What identifies my agent in Hunch's counters?
X-Hunch-Agent header on HTTP calls. Hunch counts events by client name only, with no texts, keys or addresses.Try it in one line
Connect https://hunchsheet.app/mcp/try from any MCP client with no account: ten texts a day are free, and the tools can get you a key with 100 free rows. Install lines for Claude Code, Cursor, Codex, Gemini CLI, VS Code and more. Prices are in pricing.json.