Purpose: Argot is a probabilistic guardrail for code review that judges whether a change "belongs" to the current repository, rather than judging whether the code is correct. It statistically learns rules from the repository's own Git history, flags code that deviates from existing patterns, and leaves it to humans to decide whether to accept it.
Main capabilities: It can identify issues such as unfamiliar symbols, duplicate implementations, misplaced code locations, layering violations, and test completeness; rules are learned from history and can also be extended under .argot/rules/ using TOML manifests plus sandboxed Rhai scripts, without recompiling.
Usage: First run argot audit, with no need for pre-fitting or configuration; it fits a historical baseline in a temporary worktree and evaluates the net diff from base to HEAD, without touching the working tree. Windows install command: powershell -c "irm https://github.com/get-tmonier/argot/releases/latest/download/argot-installer.ps1 | iex"; you can also use npm install -g @tmonier/argot. If the audit results are useful, then fit the current repository and score the changes to be reviewed; commit the generated argot.toml and .argot/ snapshot, merge them into the target branch of future PRs, and then add the CI workflow. CI only reads the baseline branch snapshot and does not perform fitting, so the initial snapshot PR must be separate from the CI workflow PR.
Applicable scenarios: local CLI checks, agents such as Claude Code and MCP, and CI and pre-commit integration.
Limitations: It requires usable Git history and supported source code; there is no fixed runtime commitment. Public metrics are measured separately for each detector, not the product's overall accuracy; combined briefings and runtime on ordinary repositories have not yet been publicly measured. Repositories with too shallow history, generated or vendored repositories may fail to produce useful models; it is least reliable for wrong choices that use entirely familiar vocabulary, obfuscatory wording, and code outside the selected scope. Analysis is fully local and offline, the embedding model is compiled into the binary, there is no default telemetry and no source code upload; setting ARGOT_OFFLINE=1 disables network access. The local core is free, MIT open source, and requires no account or cloud service.
