Local · no LLM · no telemetry

Lint the rules you never wrote down.

argot learns your repo’s voice from its own git history, then flags the hunks that don’t sound like anyone on your team wrote them. No model. No cloud. No GPU.

Read the docs → ★ Star on GitHub

MIT · macOS & Linux · fits in seconds, scores in milliseconds

The second question

Type checkers ask if it compiles. argot asks if it’s yours.

Linters and type checkers answer “is this valid?” They can’t answer “is this how this team writes things?” That used to live in code review — until an LLM could bury it under a hundred clean, type-correct PRs in an afternoon. argot is the layer that asks it back.

  • mypystrict — passes
  • rufflint-clean
  • argotflags one line
routers/users.py 
@router.get("/{user_id}", response_model=UserResponse)
async def get_user(user_id: int, db=Depends(get_db)) -> UserResponse:
    user = db.get(user_id)
    if user is None:
        raise ValueError(f"User {user_id} not found")
    return user
$ argot check 
argot check · 1 hunk above threshold (1 foreign)
note: argot is a probabilistic style linter — verify before action.

routers/users.py
  ●  L11-L14   8.21  foreign   · workdir · wrong exception type (bpe)
     ↳ ValueError (0×) — repo raises HTTPException (214×)
  11 │     user = db.get(user_id)
  12 │     if user is None:
  13 │         raise ValueError(f"User {user_id} not found")
  14 │     return user

Decorators, Depends, the typed return — all idiomatic FastAPI. The one break is a bare ValueError where this repo always raises HTTPException. mypy is happy. The linter has nothing. argot flags the line.

What it catches

Technically fine. Socially wrong.

argot does not replace ESLint, ruff, or your type checker. It catches the things they can’t articulate — the patterns your team agreed on by repetition, never by writing them down.

⌘

LLM paste-through

A block whose style diverges sharply from the surrounding file — fluent in the average voice of every public repo, not yours.

↯

Convention drift

Error handling, logging, or control-flow shapes that don’t match how the rest of the codebase does it.

⊘

Foreign paradigm

Class-based OOP dropped into a functional codebase. A sync def on a hot async path. The wrong import for the job.

◐

Stylistic outlier

Code that’s correct, typed, and lint-clean — but doesn’t sound like anyone on this team wrote it.

How it stays honest

Two frequency tables. No neural network.

argot builds two token distributions — one from your repo, one from a generic open-source baseline — and flags hunks far more likely under the generic one. That is the whole model. It fits on CPU in seconds and ships its threshold per repo.

Nothing leaves your machine

No GPU, no cloud, no telemetry. The model is two frequency tables and a max log-ratio — it fits in seconds and scores in milliseconds.

Calibrated per repo

The threshold is set from your own code, so “normal” means normal here — not the average of every public repo a model trained on.

Language-aware, not language-locked

A tree-sitter tokenizer parses partial, invalid hunks. Python and TypeScript out of the box; mixed monorepos get one threshold per language.

Evidence, not vibes

Every flag names the tokens that carried the score, how often they appear in your repo, and what’s common here instead.

Why argot

Reads like a linter. Thinks like a reviewer.

01

Drops into CI

argot check runs on every commit, groups hits by file, and exits non-zero when something diverges. Wire it in like ESLint.

02

Incremental, not a rewrite

Point it at a repo, run extract → fit once, then check forever. No annotations, no config to get started.

03

Per-hunk evidence

Each hit shows the offending tokens with their repo attestation — startedAt (0×) vs use (88×) — and the repo’s typical vocabulary instead.

04

Severity you can tune

unusual · suspicious · foreign, relative to the calibrated threshold. Filter the noise with --min-severity.

05

Per-language calibration

A Python + TypeScript monorepo gets one threshold per language, dispatched by file extension. No single distribution to dominate the other.

06

Honest about itself

Public benchmarks, a 35-doc research log, and a probabilistic-linter disclaimer printed in every run. Verify before you act.

Add the layer your CI is missing.

argot is MIT and alpha. Calibrate it on your repo in a couple of minutes, then see what it flags.

Get started → View on GitHub