Most AI coding tools are a single model guessing at your codebase. GitPilot deploys four specialized agents — Explorer, Planner, Coder, Reviewer — that collaborate on every task. They read your repo, draft a plan, write the code, run your tests, and wait for your approval before touching a single file. Any LLM. Free and local with Ollama. Enterprise-ready.
Four specialized AI agents share context, divide work, and check each other — the same pattern elite engineering teams use, automated at machine speed.
Every task starts with a readable, step-by-step plan. Review it, edit it, or reject it — no surprise commits.
Inline diffs in VS Code and the web app. One click to apply, one click to revert — nothing touches disk without your consent.
OpenAI, Claude, Ollama, Watsonx and OllaBridge out of the box. Swap providers in settings — no code change.
Auto-detects pytest, jest, vitest, go test and cargo. GitPilot runs them after every change and self-corrects on failure.
AI-written commit messages, PR descriptions and summaries — stitched together from the plan and the diff, not hallucinated.
Run the whole stack on your laptop with Ollama. No telemetry, no data leaves your machine unless you tell it to.
Each agent is a specialist. Together they do what no single model can: understand your repo deeply, plan safely, execute precisely, and catch their own mistakes.
Reads your full repo: file tree, git history, tests, dependencies, README. Builds the context the other agents need to make real decisions — not guesses.
Drafts a safe, step-by-step plan with line-level diffs. Surfaces risks and trade-offs before a single file is touched. You review and approve.
Edits files, runs your test suite, self-corrects on failure. Iterates until the build passes. Every risky action waits for your explicit OK.
Validates the output against the original plan, re-runs tests, and writes a commit message and PR summary ready to ship.
Pick the surface you live in. All three share the same sessions, the same settings and the same brand.
Fastest path. Python 3.11 or 3.12.
pip install gitcopilot
gitpilot serveThen open http://localhost:8000. The PyPI package is gitcopilot (the name gitpilot was taken), but the command you run is gitpilot.
The recommended experience for day-to-day coding.
The extension talks to the same backend as the web app, so your sessions, approvals and settings stay in sync.
One command, backend + frontend.
git clone https://github.com/ruslanmv/gitpilot.git
cd gitpilot
docker compose upBackend on :8000, web UI on :3000.
For contributors and power users.
git clone https://github.com/ruslanmv/gitpilot.git
cd gitpilot
make install
make runInstalls Python deps via uv and frontend deps via npm, then boots the backend + Vite dev server.
GitPilot ships with adapters for the five providers teams actually deploy in production. Change providers in settings — no code change, no restart.
Open source. Apache 2.0. Four agents, any LLM, you approve every change. Your repo, your rules.