Git commit & push in one command. With AI-generated commit messages.

Commit messages stream in token-by-token with a spinner. Diff stats, messages, and costs are color-coded. Set NO_COLOR=1 to disable colors.

go install github.com/rasalas/yeet@latest

Or build from source:

git clone https://github.com/rasalas/yeet.git
cd yeet
go build -o yeet .

yeet                     # AI generates the commit message
yeet fix typo in readme  # Use your own message
yeet -m "config"         # -m flag for words that collide with subcommands
yeet -l                  # Local commit only (skip push)

What happens:

1. Checks for staged changes — if none, runs git add --all
2. Shows diff stat (insertions in green, deletions in red)
3. Generates commit message (AI with streaming) or uses your message
4. You review — Enter to commit, e to edit inline, E to open $EDITOR, Esc to cancel
5. git commit + git push (or only commit with -l)

Selective staging works too:

git add src/auth.go
yeet                     # Only commits auth.go

Pressing Escape cancels safely — if yeet auto-staged, it unstages. If you staged manually, your staging is preserved.

yeet config

Opens a TUI where you can select your AI provider, set models, and manage keys.

Builtin providers:

Well-known providers (OpenAI-compatible API):

yeet auth set anthropic

Keys are stored in the OS keyring (macOS Keychain, Windows Credential Manager, Linux Secret Service). Environment variables (ANTHROPIC_API_KEY, OPENAI_API_KEY, etc.) are used as fallback.

Codex/Claude ACP providers use their own CLI login/config and are not managed through yeet auth.

If you have keys in environment variables or OpenCode's auth.json, import them into the keyring:

yeet auth import           # Import all available keys
yeet auth import groq      # Import a specific provider

The default provider is auto. It tries local/native providers first, then falls back to the cheapest available API-key provider by input token cost.

Default order:

1. codex via ACP, if the adapter command is available
2. ollama, if the configured Ollama server is reachable
3. claude via ACP, if the adapter command is available
4. API-key providers, sorted by input token cost

Codex/Claude subscription quota or remaining "volume" is not exposed through ACP, so yeet uses this fixed order instead of trying to guess it. Override it with:

[auto]
order = ["codex", "ollama", "claude", "api"]

Use api as a placeholder for all available API-key providers sorted by input-token cost. If you omit api, yeet will not fall back to API-key providers.

Config file: ~/.config/yeet/config.toml

provider = "auto"

[anthropic]
model = "claude-haiku-4-5-20251001"

[openai]
model = "gpt-4o-mini"

[ollama]
model = "llama3"
url = "http://localhost:11434"

codex and claude use local Agent Client Protocol adapters instead of yeet-managed API keys:

provider = "codex"  # or "claude"

The default adapters are launched with npx:

[custom.codex]
protocol = "acp"
model = "gpt-5.4-mini"  # optional; omit to use ~/.codex/config.toml
reasoning_effort = "low"  # optional; Codex defaults to lowest reasoning for yeet
command = "npx"
args = ["-y", "@agentclientprotocol/codex-acp@1.1.2"]

[custom.claude]
protocol = "acp"
model = "sonnet"  # optional; omit to use Claude's native config
command = "npx"
args = ["-y", "@agentclientprotocol/claude-agent-acp@0.32.0"]

These adapters use the agent's own auth, billing, and config (for example ~/.codex/config.toml or ~/.claude/). yeet does not store API keys for ACP providers. If codex-acp or claude-agent-acp is installed locally, yeet prefers the local binary over npx. To keep commit-message generation narrow, yeet sends the staged git context directly, advertises no client file-system or terminal capabilities, and automatically rejects ACP permission requests.

You can also set ACP models from yeet config with m. The picker reads the account-aware model choices advertised by the active Codex, Claude, or custom ACP adapter; bundled lists are only used as an offline/compatibility fallback. For Codex, t opens the thinking/reasoning picker (low, medium, high, xhigh). yeet defaults Codex reasoning to low so commit-message generation stays fast even if ~/.codex/config.toml uses a higher global default. Current adapters receive the selected model and reasoning effort through ACP session options; legacy Codex adapters keep their -c overrides. Leaving the model unset uses the adapter's native config.

Use yeet doctor --ai for a no-generation provider smoke test. For ACP providers this only checks adapter startup, protocol initialization, and session creation; it does not ask the model to generate text.

You can add custom providers that use the OpenAI Chat Completions format:

[custom.together]
model = "meta-llama/Llama-3-70b-chat-hf"
url = "https://api.together.xyz/v1"
env = "TOGETHER_API_KEY"

When generating a commit message, yeet sends the following to the AI:

• Diff — git diff --cached (truncated at 8000 lines)
• File status — git status --short for a quick overview of all changes
• Branch name — used as hint for commit type and scope
• Recent commits — git log --oneline -10 so the AI matches your style

The system prompt lives at ~/.config/yeet/prompt.txt and is created automatically on first run.

yeet prompt         # Edit in $EDITOR
yeet prompt show    # View current prompt
yeet prompt reset   # Restore default

yeet eval is an explicit, opt-in workflow for comparing prompt/model variants on real historical runs.

• Normal yeet commits stay fast — no extra model calls during commit.
• Run data is stored locally in SQLite: ~/.local/share/yeet/yeet.db
• No upload/telemetry by default.
• Requires sqlite3 to be available on PATH for eval commands/storage.
• More detail: docs/eval.md

Typical flow:

yeet eval plan --sample 20 --batch-size 5 --max-cost-usd 1.00
yeet eval generate --sample 20 --batch-size 5 --max-cost-usd 1.00 --prompt-file ./prompt-v2.txt
yeet eval judge --variant 3
yeet eval report --variant 3

Useful filters:

# Evaluate only runs where baseline output contains a phrase you want to remove
yeet eval generate --contains "for improved user" --sample 30 --batch-size 8

# Focus on runs where you had to edit the AI output
yeet eval generate --edited-only --sample 30 --batch-size 8