gitl

AI code review for your CI, with a risk score you can gate merges on — your key, any provider, or fully offline.

gitl (git-log-lens) reads a commit range and turns it into a structured engineering artifact: an AI review with a machine-readable risk level, a deterministic changelog, and a multi-repo activity digest — one Go binary, no server, no database.

gitl review — AI review of a commit range with a machine-readable risk level, streamed to the terminal.

Why

Most AI code reviewers are SaaS: your diff goes to someone else's server, and what you get back is a comment, not a contract. I wanted the opposite — a CLI/CI tool where the review runs on my key, my provider (or no provider at all), and produces a risk level CI can branch on. One binary covers three jobs that usually take three separate tools: a review with a risk gate, a changelog, and a multi-repo digest. There's no telemetry and no hosted key storage, and without a key it still works: a deterministic offline heuristic keeps CI runs green and free.

Related MCP server: grippy-code-review

Install

# Homebrew (macOS/Linux)
brew install akomyagin/tap/gitl

# npm — downloads the prebuilt, checksum-verified binary for your platform
npx gitl-cli review HEAD~5..HEAD        # or: npm install -g gitl-cli

# Go toolchain
go install github.com/akomyagin/gitl/cmd/gitl@latest

Or grab a signed binary from GitHub Releases (see VERIFY.md for how to verify it). Requires git in PATH; Go 1.22+ is only needed for go install / building from source.

Signed releases (cosign keyless + SLSA L3 provenance), no telemetry, BYOK — your API key never leaves your machine. Verify any release before you run it: VERIFY.md.

What it does

• gitl review <range> — AI review with a machine risk level (low|medium|high); --fail-on=high exits with code 2 so CI can gate on it; streams tokens live; on-disk response cache (plus an opt-in shared remote cache for CI); --staged reviews uncommitted changes; pr/N reviews a GitHub PR via gh.

• gitl changelog [<range>] — Keep-a-Changelog output grouped by conventional commits, deterministic by default; --ai rewrites it as release-note prose and falls back to the deterministic result without a key — it never fails.

• gitl digest [--days=N] [--repos=a,b,c] — activity summary by author/topic/file across multiple repos in parallel, with an interactive TUI viewer (--tui).

Providers (BYOK): OpenAI-compatible API, Ollama (local/self-hosted), Azure OpenAI, native Anthropic (Claude), Google Gemini — or no provider at all.

It's also an MCP server (gitl mcp) and ships CI wrappers for GitHub Actions, GitLab, Bitbucket, and Gitea — see the documentation site for all of it.

Quick start

# AI review of a commit range (streams to the terminal)
GITL_API_KEY=sk-... gitl review HEAD~5..HEAD

# no key = deterministic offline review (heuristic risk, no network)
gitl review HEAD~5..HEAD

# machine-readable + CI gate
gitl review HEAD~5..HEAD --format=json --fail-on=high   # exit 2 on high risk

# changelog since the last tag; activity digest for 14 days
gitl changelog
gitl digest --days=14

Exit codes are part of the contract: 0 — ok, 1 — tool/runtime error, 2 — the --fail-on risk gate triggered. Full command reference → documentation site.

--fail-on=high turns a high-risk range into a non-zero exit code (2) your CI can gate on.

Use cases

• Gate a PR on AI risk in CI — the GitHub Action posts a sticky review comment and fails the check above your threshold.

• Pre-commit safety net — gitl review --staged runs offline before every commit, at zero cost.

• Release notes in one command — gitl changelog --ai.

• Multi-repo standup digest — gitl digest --repos=… --tui.

See these worked end-to-end on the use cases page.

gitl digest --tui — an interactive viewer for the multi-repo activity summary.

How it compares

Feature comparison as of 2026-09; corrections welcome.

Documentation

Full usage docs, the configuration reference, and worked use cases live on the documentation site (also in Russian).

Русская версия README — README_RU.md.

Contributing

I develop this in the open, and it's a one-person project — issues, questions, and bug reports from real CI setups are the most valuable thing you can send my way. Open an issue or a PR; I read everything.

License

MIT.