Contents.

10 chapters. Read them in order if you are new; every code sample on every page is run by the test suite.

1. Getting startedread →

   Install, pick a model, and the three questions you can ask.

2. Backendsread →

   Every model gut can run on — Jev (directly, through OpenRouter, or open models on Ollaya), local NLI, local LLMs, Ollama, vLLM, OpenAI — Cascade, and writing your own.

3. Knowing when it doesn't knowread →

   lean, ask_human, stakes — how careful to be, in words. What if and match do with UNSURE.

4. Asking everything at onceread →

   @semantic and judge(): every judgment about one subject, together.

5. Asyncread →

   alikely, aclassify, arate, and the async side of each(), @semantic and judge().

6. Exact costsread →

   The cost model underneath the posture words, its formula, and the trap in it.

7. Caching and observabilityread →

   The cache, seeing every decision as it is made, and what the calls cost.

8. Command lineread →

   gut filter and gut map: judgments over lines of text from a shell, with a cost summary and a budget.

9. MCP serverread →

   gutfeel-mcp: the same judgments as tools for Claude Code, Claude Desktop, Cursor and any MCP client.

10. Honest limitationsread →

    What small models get wrong, and what gut does not do.

+Also worth knowing about

• examples/ ↗ — the same handlers running on any backend you pick from the command line, including a keyword rule turned into the first stage of a cascade.
• skills/gut/SKILL.md ↗ — the compact version, written for a coding agent. Point your agent at this rather than at the docs.
• llms.txt ↗ — the machine-readable index.

Every code block on these pages is executed by the test suite, so none of it can drift from the library.