What it does

argot compares a set of code changes with what a repository has historically looked like. It learns from the repo's own git history using statistics rather than a second language model, then flags changes that stand out: imports of dependencies the repo has never used, calls to unfamiliar APIs, new functions that duplicate existing ones, code placed in the wrong module, and imports that reverse the project's layering. It also watches for signs that tests were bent to pass, such as tests removed, skip markers added or assertions loosened alongside production changes. Teams can add their own rules with a TOML manifest and a sandboxed Rhai script. The project stresses that findings are prompts for a human to review, not proof that code is wrong. argot runs locally and offline, because the code-embedding model ships inside the binary, and the project says it has no default telemetry and does not upload source code. An audit command works without prior setup, while init and check use a learned snapshot that is committed to the repository. It can be used from the command line, as a Claude Code plugin, through MCP, as a pre-commit hook or in a GitHub Action.

Capabilities

Learns repository conventions from git history with statistics, not a second LLM
Flags foreign imports, unfamiliar calls, duplicated or misplaced functions and layering breaks
Detects tests that were removed, disabled or weakened
Runs fully local and offline with no account
Custom rules through a TOML manifest and Rhai scripts
Works from the CLI, a Claude Code plugin, MCP, pre-commit and a GitHub Action

When to use it

Reviewing pull requests that contain AI-written code against the repo's own patterns
Adding an advisory check to pre-commit or CI without sending code to a cloud service
Auditing recent changes in a repository that has usable git history