autocommit

AI-powered commit and pull request generation that runs entirely on your machine. Analyzes local git diffs using a bundled LLM and drafts structured, high-signal commit messages and PR descriptions through an interactive terminal workflow.

No API keys. No cloud. Fully local inference powered by llama.cpp.

Features

• Commit generation — analyzes staged/worktree diffs, generates conventional commit messages with interactive approve/edit/cancel flow
• PR generation — generates PR title and body, creates or updates pull requests via gh, with branch selection and optional push
• Version bump recommendations — detects manifest files (Cargo.toml, package.json, go.mod, etc.), suggests semver bumps, syncs lockfiles after applying
• Local inference — ships with llama.cpp compiled in, supports Metal (macOS), CUDA (NVIDIA), Vulkan (cross-vendor), and SYCL (Intel) GPU acceleration
• Per-repo config — caches model state and runtime settings under .git/autocommit/ for fast subsequent runs

Installation

Pre-built binaries

Download from GitHub Releases:

# macOS (Apple Silicon)
curl -L https://github.com/MicroPerceptron/autocommit/releases/latest/download/autocommit-aarch64-apple-darwin.tar.gz | tar xz
sudo mv autocommit /usr/local/bin/

# macOS (Intel)
curl -L https://github.com/MicroPerceptron/autocommit/releases/latest/download/autocommit-x86_64-apple-darwin.tar.gz | tar xz
sudo mv autocommit /usr/local/bin/

# Linux (CPU)
curl -L https://github.com/MicroPerceptron/autocommit/releases/latest/download/autocommit-x86_64-unknown-linux-gnu.tar.gz | tar xz
sudo mv autocommit /usr/local/bin/

# Linux (CUDA — NVIDIA GPUs)
curl -L https://github.com/MicroPerceptron/autocommit/releases/latest/download/autocommit-x86_64-unknown-linux-gnu-cuda.tar.gz | tar xz
sudo mv autocommit /usr/local/bin/

# Linux (Vulkan — Intel/AMD/NVIDIA GPUs, requires Vulkan drivers)
curl -L https://github.com/MicroPerceptron/autocommit/releases/latest/download/autocommit-x86_64-unknown-linux-gnu-vulkan.tar.gz | tar xz
sudo mv autocommit /usr/local/bin/

# Linux (SYCL — Intel GPUs, requires Intel oneAPI runtime)
curl -L https://github.com/MicroPerceptron/autocommit/releases/latest/download/autocommit-x86_64-unknown-linux-gnu-sycl.tar.gz | tar xz
sudo mv autocommit /usr/local/bin/

cargo binstall

cargo binstall autocommit-cli

Build from source

Requires: Rust toolchain, Git, CMake, C++17 compiler, libclang

git clone --recursive https://github.com/MicroPerceptron/autocommit.git
cd autocommit
cargo install --path crates/cli --locked --features llama-native

GPU backend selection

By default, Metal is used on macOS and CUDA is auto-detected on Linux. To build with a different GPU backend:

# Vulkan (requires Vulkan SDK + glslc)
GGML_VULKAN=ON cargo install --path crates/cli --locked --features llama-native

# SYCL / Intel oneAPI (requires oneAPI toolkit)
source /opt/intel/oneapi/setvars.sh
GGML_SYCL=ON cargo install --path crates/cli --locked --features llama-native

Only one GPU backend can be active at a time.

Quick start

# 1. Initialize — select a model and configure commit policy
autocommit init

# 2. Make some changes, then generate a commit
autocommit commit

# 3. Generate a pull request
autocommit pr --push

Usage

Commands

Command	Description
autocommit commit	Generate and create a commit from local changes
autocommit pr	Generate and create/update a pull request
autocommit init	Initialize per-repo model config and commit policy
autocommit config	View and update per-repo settings
autocommit analyze	Analyze diff and emit a structured JSON report
autocommit clean	Remove persisted KV generation cache

Run autocommit <command> --help for detailed options.

Common workflows

# Interactive commit (default)
autocommit commit

# Non-interactive / CI mode
autocommit commit --yes --no-interactive --staged

# Dry run — preview without creating a commit
autocommit commit --dry-run

# PR targeting a specific base branch
autocommit pr --base origin/main --push

# Use a specific model
autocommit commit --hf-repo ggml-org/Qwen3-4B-GGUF:Q4_K_M

Model configuration

Model settings are resolved in order:

1. CLI flags (--model-path, --hf-repo, --cache-dir, --profile)
2. Per-repo metadata saved by autocommit init
3. Runtime defaults

Supported model sources:

• Hugging Face — --hf-repo ggml-org/Qwen3-1.7B-GGUF:Q4_K_M (auto-downloaded and cached)
• Local GGUF — --model-path /path/to/model.gguf

List cached models:

autocommit init --list-cached-models

Architecture

crates/
  core/           Pure Rust analysis pipeline, dispatch, and diff processing
  cli/            Terminal application (clap, dialoguer, indicatif)
  llama-runtime/  Runtime integration layer wrapping llama-sys
  llama-sys/      Native llama.cpp bindings (CMake build + bindgen FFI)
third_party/
  llama.cpp/      Upstream submodule (compiled from source, statically linked)

Development

# Run tests
cargo test --workspace --features llama-native

# Lint
cargo clippy --workspace --features llama-native -- -D warnings

# Format
cargo fmt --all

# Quick check (no native build)
cargo check -p autocommit-cli

Troubleshooting

Problem	Solution
runtime model path is not configured	Run autocommit init or pass --hf-repo / --model-path
GPG signing failures	Run autocommit commit --configure-commit-policy to adjust signing settings
PR creation errors from gh	Ensure gh auth status succeeds and the base/head branches differ

License

MIT