01
Human-in-the-loop by design
Every commit, push, and PR message passes an explicit approval step. The AI proposes; the developer disposes. Automation never outruns intent.
A cross-platform AI Git CLI — conventional commits, semantic grouping, and PR summaries with a human always in the loop.
2025 · Developer Tooling · role: owner
● status: production·6 services·~2 min read
01 — overview
GitPilot automates the tedious parts of Git hygiene — writing good conventional commit messages and organizing changes into clean, logical commits — while keeping the developer in control through an approval-before-commit workflow. It ships as a single binary for macOS, Linux, and Windows.
02 — the problem
Writing consistent conventional commits and splitting a messy working tree into logical commits is tedious — so most developers don't, and history quality suffers.
AI dev tools routinely store API keys in plaintext config files, turning a convenience into a credential-leak vector.
03 — architecture
A single Go binary wraps the local Git workflow. Diffs flow through a provider-agnostic LLM layer; every AI suggestion passes a human approval gate before touching the repository; credentials resolve from the OS keychain first.
04 — engineering decisions
01
Every commit, push, and PR message passes an explicit approval step. The AI proposes; the developer disposes. Automation never outruns intent.
02
API keys resolve from the OS keychain — macOS Keychain, Linux libsecret, Windows Credential Manager — with a chmod-0600 config fallback. CI/CD uses environment variables only.
03
Beyond file-wise splitting, the LLM clusters related changes across files into logical commits with conventional prefixes — turning one messy working tree into a readable history.
04
GoReleaser ships SHA256-verified binaries via Homebrew tap, go install, and one-line installers for every platform.
Explain this project
A single Go binary that wraps the local Git workflow: it feeds diffs to a provider-agnostic LLM layer (Groq, OpenRouter, OpenAI, Claude, Ollama), generates conventional-commit messages, and clusters related changes across files into semantic commit groups. Every AI suggestion passes an explicit human approval gate before any commit, push, or PR. Credentials resolve from the OS keychain first (macOS Keychain, Linux libsecret, Windows Credential Manager) with a chmod-0600 fallback, and releases ship SHA256-verified via GoReleaser and a Homebrew tap.
05 — features
06 — stack
The repository has the code, the commits, and the decisions in context — or reach out and I'll walk you through it.