AI-powered Git commit message generator. Analyzes your staged changes and generates Conventional Commits formatted messages using any OpenAI-compatible LLM or Claude Code.
brew tap lifedever/tap
brew install ai-commitNewer Homebrew versions require a one-time approval for third-party taps. If you see an "untrusted tap" error, run
brew trust lifedever/tapand retry.
curl -fsSL https://raw.githubusercontent.com/lifedever/ai-commit/main/install.sh | bashRequires: Git, Node.js >= 22.19, npm
Set your API key (add to ~/.bashrc or ~/.zshrc for persistence):
export AI_COMMIT_API_KEY="sk-your-api-key"By default, it uses DeepSeek. To use other providers:
# OpenAIexport AI_COMMIT_API_URL="https://api.openai.com/v1/chat/completions"export AI_COMMIT_MODEL="gpt-4o-mini"# Volcengineexport AI_COMMIT_API_URL="https://ark.cn-beijing.volces.com/api/v3/chat/completions"export AI_COMMIT_MODEL="deepseek-v3-2-251201"# Local Ollamaexport AI_COMMIT_API_KEY="ollama"export AI_COMMIT_API_URL="http://localhost:11434/v1/chat/completions"export AI_COMMIT_MODEL="qwen2.5"Any OpenAI API-compatible service works out of the box.
If you have Claude Code installed, you can use it as the provider. Claude Code can read your source files for better context understanding, producing higher-quality commit messages.
export AI_COMMIT_PROVIDER="claude"# No API key needed — Claude Code manages its own authenticationLarge diffs: When staged changes exceed 15K characters, ai-commit automatically truncates the diff and passes it directly to Claude instead of letting Claude read files itself. This prevents timeouts on very large changesets.
# Stage your changes first
git add .# Generate commit message (interactive)
ai-commit
# Auto-commit without confirmation
ai-commit -y
# Stage all tracked changes first (like git commit -a)
ai-commit -a
# Preview only, don't commit
ai-commit --dry-run
# Use Chinese for commit message
ai-commit -l zh
# Use a specific model
ai-commit -m gpt-4o-mini
# Add emoji to commit message (e.g. ✨ feat: add feature)
ai-commit --emoji
# Use Claude Code as provider
ai-commit -p claude
# Or set it permanentlyexport AI_COMMIT_PROVIDER="claude"
ai-commitIf nothing is staged but tracked files have unstaged changes, ai-commit lists them and asks whether to stage them (y/N). Untracked files are never staged automatically —
git addthem explicitly. With-y, nothing is ever staged implicitly.
| Option | Description |
|---|---|
-V, --version | Show version number |
-a, --all | Stage all tracked changes first (like git commit -a; untracked files excluded) |
-y, --yes | Auto-commit without confirmation |
-l, --language <lang> | Set commit message language (en / zh) |
-m, --model <model> | Use a specific model |
-e, --emoji | Add emoji to commit message |
-p, --provider <provider> | LLM provider (openai / claude) |
-d, --dry-run | Preview message only, don't commit |
--update | Update ai-commit to the latest version |
--uninstall | Uninstall ai-commit |
-h, --help | Show help |
After each generation, a dim summary line is displayed showing:
──────────────────────────────────────────────────
feat(auth): add JWT token refresh mechanism
──────────────────────────────────────────────────
provider: openai · model: deepseek-chat · tokens: 156 · time: 2.8s
Each run automatically checks for new versions (non-blocking, cached for 24 hours). If a newer version is available:
💡 New version v1.4.0 available, run ai-commit --update to update
git config --global alias.ac '!ai-commit'# Then use:
git acai-commit --updateai-commit --uninstall| Variable | Description | Default |
|---|---|---|
AI_COMMIT_PROVIDER | LLM provider (openai / claude) | openai |
AI_COMMIT_API_KEY | Required for openai provider. Your LLM API key | - |
AI_COMMIT_API_URL | API endpoint | https://api.deepseek.com/v1/chat/completions |
AI_COMMIT_MODEL | Model name | deepseek-chat |
AI_COMMIT_LANGUAGE | Message language and CLI language (en / zh) | en |
AI_COMMIT_MAX_TOKENS | Max tokens for generation | 500 |
AI_COMMIT_EMOJI | Always add emoji (true / false) | false |
OpenAI provider supports system proxy via standard environment variables:
export HTTPS_PROXY="http://127.0.0.1:7890"# orexport HTTP_PROXY="http://127.0.0.1:7890"Both HTTPS_PROXY / https_proxy and HTTP_PROXY / http_proxy are recognized. NO_PROXY / no_proxy is also supported to bypass proxy for specific hosts.
See CHANGELOG.md for release history.
MIT