gut MCP server

Judgment calls as one line of Python — built for Jev, and running on any small model.

Try it in your browser → The site runs gut's local model in the page: no key, no server.

Your code keeps running into questions that aren't logic: Is this comment spam? Which team owns this ticket? How urgent is it? Is the agent's task done? Until now there were three answers:

• Regex and keyword rules — free and instant, and wrong the moment someone phrases it differently.
• A frontier LLM — understands anything, at seconds and cents a call, with prose to parse.
• Train a classifier — cheap to run, once you have the labelled data, the pipeline and the week.

There is a fourth: a small model made for exactly these questions. TypeSafe AI's Jev answers typed questions directly — a probability for yes, a distribution over options, a score on a scale — with nothing to generate or parse, billed on input only. gut is built around it, and makes it a line of code:

import gut

gut.configure(backend=gut.JevBackend())   # or just set TYPESAFE_API_KEY

if gut.likely(comment, "is spam"):
    hide(comment)

No prompt, no parsing, no threshold — and no model named at the call site.

Three questions

gut.likely(ticket, "is a bug report")                      # yes / no
gut.classify(ticket, Team)                                 # which one — an Enum
gut.rate(ticket, ["can wait", "this week", "right now"])   # how much

It knows when it doesn't know

A regex never hesitates, and neither does an LLM. gut can:

match gut.likely(email, "the customer threatens to cancel", ask_human=True):
    case gut.YES:    escalate(email)
    case gut.NO:     auto_reply(email)
    case gut.UNSURE: send_to_a_person(email)

Say how careful to be in words — lean="yes", stakes="high" — and gut works out the thresholds.

A thousand subjects, one line

spam = gut.each(comments).likely("is spam")     # one decision per comment, in order
teams = gut.each(tickets).classify(Team)

Jev gets concurrent requests, a local model batched passes, and nothing already cached is asked twice. @gut.semantic does the same for several questions about one subject.

Jev first, any model

Jev is the model gut is designed around. It is not the only one: the model is configuration, and the same line runs unchanged on any of these.

gut.configure(backend=gut.JevBackend())                            # TypeSafe AI's Jev
gut.configure(backend=gut.JevBackend.openrouter())                 # Jev, through OpenRouter
gut.configure(backend=gut.JevBackend.ollaya("winnow:e4b"))         # open decision model, Ollaya
gut.configure(backend=gut.ZeroShotBackend())                       # NLI model, on your CPU
gut.configure(backend=gut.TransformersBackend("Qwen/Qwen3-0.6B"))  # small LLM, on your machine
gut.configure(backend=gut.OpenAICompatibleBackend(                 # Ollama, vLLM, llama.cpp
    "qwen2.5:1.5b", base_url="http://localhost:11434/v1"))
gut.configure(backend=gut.OpenAICompatibleBackend("gpt-4.1-nano")) # OpenAI

Or several at once. Cascade asks the cheapest model first and passes on only what it is unsure of:

Shortened. The full README is on GitHub.