argot

Lint the rules you never wrote down.

AI writes the code. argot harnesses it with the one thing that can’t hallucinate: your repo’s own history .

Statistics, not a second LLM. 100% local. It surfaces the divergence — you decide what to accept.

argot.tmonier.com · Documentation · Evidence · Research log

🎬 Watch the 45-second launch film

Start with an audit

argot audit needs no prior Argot fit or configuration. It fits a historical base in a temporary worktree, then evaluates the surviving base-to-HEAD net diff. Your working tree is left untouched. It is a review prompt—not a census of who wrote code, or proof that a finding is a defect.

# macOS / Linux

curl --proto '=https' --tlsv1.2 -LsSf https://github.com/get-tmonier/argot/releases/latest/download/argot-installer.sh | sh

cd your-repository argot audit

Windows: powershell -c "irm https://github.com/get-tmonier/argot/releases/latest/download/argot-installer.ps1 | iex". The npm package is also available as npm install -g @tmonier/argot.

Audit needs usable Git history and supported source. It has no fixed runtime promise. It runs fully offline — the code-embedding model behind the semantic findings ships inside the binary. See Getting started for install and fit details.

If the audit gives you a useful lead, fit the current repository and score the changes you intend to review:

argot init

argot check

Review and commit the generated argot.toml and .argot/ fit snapshot, then merge it into the branch future PRs target before adding a CI workflow. Local tools and CI then use the same learned baseline. CI only reads the base branch snapshot; it never fits, so the initial snapshot PR must be separate from the CI-workflow PR. argot status later recommends a local fit-and-commit refresh only when accepted source, function, or layout surfaces have materially changed. Commit count and age are not refresh triggers by default; [fit] refresh-after is available only as an explicit team backstop. The argot-refresh skill re-audits exclusions, structural paths, and mutes before fitting, so a reorganized repository does not blindly relearn old scope.

flowchart LR

A["argot init learn locally"] --> B["review + commit argot.toml · .argot/"] B --> C["local tools + CI read one baseline"] C --> D{"material accepted drift?"} D -- no --> C D -- yes --> E["argot-refresh review scope · fit locally"] E --> B

The embedding model itself ships inside the binary. Git stores only the repository-specific learned snapshot—typically a few MB to a few tens of MB—so every clone can reproduce the check without retraining or operating a service.

check reports patterns worth reviewing on the selected changeset; a clean result does not prove the change correct or fully idiomatic. Read the Audit, Init and Fit, and Check guides for the exact contracts.

What it surfaces

Type checkers ask if it compiles. argot asks if it’s yours. A clean, type-correct, well-reviewed pull request can still be foreign to the repository it lands in. These are the rules argot ships, every one of them learned from your own history rather than configured by hand:

Repositories add their own on top — a TOML manifest plus a sandboxed Rhai script under .argot/rules/, with working ones to copy in examples/rules/. No recompilation.

argot is a probabilistic review guardrail, not a correctness oracle. Each finding carries repository evidence. Treat it as a prompt to inspect and make the human decision explicit—never as proof that the code is wrong.

Choose how to run it

The CLI is the complete, explicit changeset check. Other routes have narrower triggers and coverage; none provides a universal acceptance-time check.

Canonical setup and host details: Claude Code, other agents and MCP, and CI and pre-commit.

Evidence and limits

Current public measurements are detector-specific, not a product-wide accuracy or combined-brief claim. The approved claim manifest records:

• visible foreign-symbol fixtures: 620/637 — 97.3% across 36 corpora and 12 languages;
• reinvention fixtures: 545/584 — 93.3% across 31 corpora and 11 languages;
• placement transplants: 12,899/13,456 — 95.9% across the 22 evaluable corpora and 11 languages (the other nine abstain because their layouts have no separable architecture);
• layering fixtures: 264/272 — 97.1% across 25 corpora and 12 languages;
• test-integrity fixtures: 154/164 — 93.9% across 23 corpora and 12 languages.

A catch rate means little without the noise it costs, so both are published. On the same 36 corpora, the voice detectors flag 0.25% of ordinary accepted edits — and 0.00% of the hunks in newly added files, where a repository has the least to say about what belongs.

Each number has a distinct corpus, denominator, and qualifier. The combined briefing/noise result and ordinary-repository timing are not yet measured public claims. See the benchmark methodology and sources.

Argot ships adapters for 12 languages. The five tested release targets are macOS arm64/x64, Linux x64/arm64, and Windows x64. The local analysis path uses statistical, graph, scripted, and embedding evidence; no generative or opinion-forming model decides a finding.

Fit health matters. A repository with shallow, generated, vendored, or otherwise unsuitable history