Judgment calls as one line of Python.

Between a regex and a frontier model, there is a lot of room.

Most judgment calls in a codebase are small: is this spam, does this mention a refund, is this urgent. A regex is too blunt for them. A frontier LLM is far more than they need. gut lives in the middle.

Three questions cover almost every judgment call.

Is it? Which one? How much? Each is a function, each returns a typed answer, and none of them asks you to parse anything.

You never write a threshold. You say how careful to be.

Three words set the posture: lean, stakes and ask_human. gut works out where the boundaries go. Drag the needle, change the words, and watch the hatched band move.

The real model, running on your machine.

This is gut's ZeroShotBackend model, loaded into this page. No key, no server. Pick a preset or write your own.

This is the 70M model running on your machine. Jev, and larger models, are sharper; the code does not change.

One line of code. Six places to run it.

gut is built around TypeSafe's Jev, a model that answers typed questions directly, with probabilities. It is also model-agnostic: the same line runs on a laptop CPU, an open decision model on Ollaya, a small local LLM or any OpenAI-compatible server.

1. JevBackendgutfeel[jev]TypeSafe's Jev. Answers typed questions directly with probabilities. Set TYPESAFE_API_KEY and go, or reach it through OpenRouter with JevBackend.openrouter(). JevBackend.ollaya() asks an open decision model on your own machine instead.
2. ZeroShotBackendgutfeel[local]A 70M NLI model on a laptop CPU, about 0.1 s a question. Free, offline, and hard to talk round.
3. TransformersBackendgutfeel[local]A small local LLM such as Qwen3-0.6B, for when you want a language model without a server.
4. OpenAICompatibleBackendgutfeelAnything that speaks the OpenAI API: Ollama, vLLM, llama.cpp, or OpenAI itself.
5. CascadegutfeelCheap model first. Only the answers it is unsure of go on to the next.
6. FakeBackendgutfeelFor tests. No model, answers you script.

gut.configure(backend=gut.Cascade(
    gut.ZeroShotBackend(),   # free and local: settles the obvious
    gut.JevBackend(),        # sees only what the first could not
))

Many subjects, one call. And async twins.

gut.each() batches a question over a list. Every function has an async twin: alikely, aclassify, arate.

spam = gut.each(comments).likely("is spam")
# many subjects, one call

decision = await gut.alikely(email, "...")
# alikely, aclassify, arate

Evidence, from the examples folder.

Real output from scripts in the repository. Printed as it came out, including the miss.

Your coding agent can learn it too.

The big model thinks; the small one decides, fast. From a shell, gut filter judges a thousand files, commits or log lines in one command, so an agent never has to read them all. The skill teaches it when to reach for that, and how to write gut in code.

# teach your coding agent the skill
npx skills add Kungie/gut --skill gut
 
# then it can judge at scale, with a budget
git ls-files | gut filter "retries failed requests" \
  --read-files --max-cost 0.50

Let the big model delegate the small calls.

gutfeel-mcp hands any MCP client, such as Claude Code, Claude Desktop or Cursor, gut's judgments as tools, so the cheap calls go to a small model instead of a big one.

• likelyYES, NO or UNSURE on one claim, with its probability.
• classifyOne of the options you give it.
• rateA rung on the scale you give it.
• eachOne question over up to a thousand texts.

claude mcp add gut --env TYPESAFE_API_KEY=your-key -- uvx --from "gutfeel[mcp]" gutfeel-mcp

{
  "mcpServers": {
    "gut": {
      "command": "uvx",
      "args": ["--from", "gutfeel[mcp]", "gutfeel-mcp"],
      "env": { "TYPESAFE_API_KEY": "your-key" }
    }
  }
}

Contents.

Ten chapters, read in order if you are new. Every code sample in them is run by the test suite.

1. Getting startedread →
2. Backendsread →
3. Knowing when it doesn't knowread →
4. Asking everything at onceread →
5. Asyncread →
6. Exact costsread →
7. Caching and observabilityread →
8. Command lineread →
9. MCP serverread →
10. Honest limitationsread →

Small models are small.