argotdocs
BenchmarksGitHub ↗

Getting started

Install argot, audit accepted history, then fit a voice you can use in the daily review loop.

Argot learns patterns from a repository’s Git history and surfaces changes that look foreign to that repository. It is a probabilistic review aid: a finding is a prompt to inspect evidence, not an instruction to reject a change.

Install

argot is a single static binary — no Python, no Node, no runtime to install, and nothing to fetch afterwards. The code-embedding model the semantic layer uses is compiled into the binary, so the first run works on an air-gapped machine and nothing ever leaves it.

macOS / Linux — package-manager-free installer:

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

Windows — the PowerShell equivalent:

powershell -c "irm https://github.com/get-tmonier/argot/releases/latest/download/argot-installer.ps1 | iex"

Any platform — prefer a package manager?

npm install -g @tmonier/argot

All three download the prebuilt argot binary for your platform (macOS arm64/Intel, Linux x64/arm64, Windows x64). Everything runs locally — no API key, no account, nothing leaves your machine.

Start with accepted history

Run an audit before configuring anything. It creates a temporary worktree, fits a historical base, and reports what would have prompted review in the surviving base-to-head changes. It does not modify your tree and exits 0 when it completes.

cd your-repo
argot audit

Read the result in the Audit guide, then decide whether to set up a local voice.

Fit, then review a change

init is the portable setup command. It writes a commented argot.toml when one is absent, gitignores personal argot.local.toml, fits the voice, and prints its health. Create that shared configuration before reviewing suggestions: --suggest only reports evidence and never edits or fits configuration.

argot init
argot init --suggest
# Review and edit argot.toml [exclude].paths if appropriate.
argot init
argot check

Use a clean default-branch checkout for a manual fit whenever possible. A dirty tree or unmerged source commits on a feature branch are warned about because manual fitting learns files as they are on disk. The Init and Fit guide explains the artifacts and refresh behavior.

The model to keep in mind is deliberately small:

  1. Learn locally. init builds the repository-specific voice and indexes; nothing is uploaded.
  2. Review, commit, and merge the baseline. Commit argot.toml and the generated .argot/ snapshot so every clone sees the same learned repository; merge that baseline into the branch future pull requests target before enabling CI. The embedding model itself is already in the binary.
  3. Check everywhere. Local tools and CI read that snapshot. Pull requests are judged against the base branch copy, so add CI only after that branch carries the baseline; CI never fits.
  4. Refresh only after material accepted drift. status, checks, MCP, and CI explain when the learned source/function/layout surface has moved enough. Run argot-refresh locally to review corpus paths and mutes before fitting and recommitting; there is no default time or commit cadence.

check scores the change you select. Its exit code is command-specific: 0 means no error-severity findings, 1 means review findings, and 2 is a setup or usage error. Read the Check guide before configuring any local or CI response to those results.

Set it up properly, once

Scoping decides everything downstream: argot is only as good as the judgment of which code is your voice. Set up with your agent is one copy-pasteable prompt that walks any coding agent through it — audit, scope, fit, verify, tune, and wire local and CI together. If your agent supports skills, npx skills add get-tmonier/argot gives you the same flow as argot-setup.

Where to next