git-lrc

Purpose

git-lrc is a free, lightweight AI code review tool that hooks into the git commit process and automatically runs AI review on the code diff before each commit. Its goal is to catch security vulnerabilities, reliability issues, and hidden technical debt that may be introduced by AI-generated code or manual changes as early as possible, before the code enters permanent Git history.

Main Capabilities

• Automatic review at commit time: Through Git hooks, it automatically scans the diff on each commit, with no manual trigger required.
• Risk classification and pattern detection: Covers 10 risk categories and more than 100 failure patterns, including security (such as credential leaks), reliability (such as error handling), and cost (such as expensive cloud calls).
• Issue Navigator: Organizes review results into a structured, filterable view, supports filtering by severity (Critical, Warning, Info), category/subcategory, type, and area, and displays real-time issue counts.
• Risk scoring view: Sorts each code block (hunk) by two independent scores from 0-100: "Blast Radius" and "Review Priority," helping prioritize the changes with the greatest impact. Scoring runs locally based on a codebase knowledge graph, in parallel with server-side AI review.
• Review summary slide: Each completed review generates a brief slide summary outlining what changed, why, and the flagged risks.
• Git log tracking: The git log message of each commit includes an appended review status line recording the iteration count (iter) and coverage, letting teams know which commits were reviewed, vouched for, or skipped.
• Agent mode:--agent-mode Runs AI review non-interactively and outputs findings in JSON format, suitable for coding agents (such as Claude Code, Cursor) or CI environments.
• claude-lrc: Bundled with git-lrc installation, it provides slash commands integrated with Claude Code, allowing review, vouching, and skipping directly in chat.
• Bring your own AI connector (BYOK): In addition to the default Gemini setup, it supports bringing your own API keys, including Anthropic-compatible APIs, Atlas Cloud, and more, and allows reordering connector priority in the UI.
• Repository rules: You can write rules in the .lrc/ directory at the repository root (rules/*.md, limit 3000 characters), ignore patterns (ignore), etc., so the reviewer understands team preferences and constraints.

How to Use

1. Installation: Supports platforms such as Windows (PowerShell), and after installation sets up hooks globally.
2. One-time configuration: Requires two API keys—the LiveReview API key (log in via Hexmos) and a free Gemini API key (obtained from Google AI Studio). It takes about 1 minute to complete, takes effect machine-wide, and afterward all Git repositories will trigger review on commit without per-repository configuration.
3. Review workflow:
   • Automatic mode: automatic review on commit.
   • Manual mode: run before committing git lrc review. After review, you can view GitHub-style diffs, inline AI comments, severity badges, review summaries, etc. in the browser.
   • You can copy issues and feed them back to an AI agent for fixes, repeating git add . → git lrc review until satisfied.
   • Use git lrc review --vouch to commit after manual vouching; or use --skip to skip review.
   • Agent mode:--agent-mode Output JSON, after successful review record the agent-reviewed status, and submission can continue.

Applicable scenarios

• Teams using AI agents to generate large amounts of code and needing to prevent hidden damage (such as deleting logic, loosening constraints, introducing expensive cloud calls, leaking credentials).
• Developers who want to catch problems at the commit stage rather than the PR stage.
• Engineering teams that need a unified review process across repositories and editors.
• Teams that want to provide institutional memory for onboarding new members, post-incident reviews, or tracing historical changes.

Limitations

• The free quota is 30,000 lines of code (LOC) per month; the premium version starts at $32/100,000 LOC and is billed by workload rather than by number of people.
• API keys need to be configured (LiveReview and Gemini or a self-provided connector).
• The risk scoring engine may take several minutes to index a large repository for the first time, and subsequent incremental refreshes take about one second; scoring is disabled when the engine is not used.
• policy/ The directory is in TODO status and is not yet enforced.
• Repository rules rules/ The total character limit is 3000.
• Some features (such as Windows alternative installation and complete setup steps) require reference to the official Get Started documentation.