I got tired of context-switching between coding and git operations, so I built vibe-git - an MCP server that automates the entire git workflow when AI assistants make code changes.

Here's how it works:

- You say "Let's start a vibe session" → AI calls start_vibing(), server creates a branch and starts a file watcher

- The watcher detects when AI edits files and auto-commits changes (every second)

- You say "Let's stop vibing" → AI calls stop_vibing(), server squashes commits, rebases on latest main, and creates a PR

The "vibe" philosophy: you speak naturally about features, AI codes, file watcher handles git seamlessly. Plus you never lose work if/when the agent goes rogue since every change is committed as it happens.

I've been dogfooding this by developing vibe-git using Claude Code + vibe-git.

The current stack is implemented in Python + FastMCP + beartype + plum + mutmut.

My experience so far is it's much easier to stay in the flow when you don't have to context switch to committing / managing git and I've found I micromanage the agent less and instead focus on iterating towards a good looking PR.