[Open Source] I built a full Git MCP server in Go that doesn't just wrap bash. It uses tree-sitter, handles real plumbing (write-tree), and runs 100% locally.

I was tired of watching LLM agents fail at basic Git operations. Standard integrations pass raw text, hang on pagers, or scream because they can't parse unstructured ⁠git diff⁠ outputs. **git-courer** is a full Model Context Protocol (MCP) server written in Go that treats Git properly. No bash spawning, no unstructured text to parse. Everything communicates via structured JSON. Here is an actual commit message it generated completely locally: fix: fix mcp server connection handling WHY The previous implementation lacked proper error handling for connection failures in the MCP server, leading to unhandled panics or silent failures when the local LLM backend was unreachable. WHAT \* Added connection timeout logic to the local client calls. \* Implemented retry mechanisms with exponential backoff for transient backend errors. The Architecture & Tool Pack **Read Tools (status, diff, history, blame):** Completely structured JSON and fully paginated. A single ⁠status⁠ call replaces over 5 standard Git commands for the agent. **Write Tools (commit, merge, rebase, branch, stash, stage, sync...):** Every single mutation auto-creates a backup before executing. If the LLM messes up, a ⁠RESTORE⁠ command brings you back exactly where you were. **Safety Model:** Destructive operations (hard resets, force pushes, branch deletions) require an explicit ⁠confirmed=true⁠ gate. The agent is forced to ask you first. ⁠dry\_run=true⁠ is also available for peace of mind. The Semantic Annotator (Why it's different) Instead of just feeding raw code to the LLM, git-courer uses ⁠go-enry⁠ + ⁠go-tree-sitter⁠ to parse the AST and tag every hunk semantically before the LLM even sees it. It detects tags like ⁠NEW\_FUNC⁠, ⁠MOD\_SIG⁠, ⁠MOD\_BODY⁠, ⁠DELETED⁠, and ⁠BREAKING\_CHANGE⁠. The commit type (⁠feat⁠, ⁠fix⁠, ⁠refactor⁠) is determined deterministically from these AST tags rather than guessed by the model. The Commit Pipeline **Atomic Commits:** One staged area = one commit. It actively prevents the agent from creating giant, messy multi-feature commits. **In-Memory Previews:** The ⁠PREVIEW⁠ tool uses ⁠write-tree⁠ to snapshot the staging area into a ⁠job\_id⁠. The working tree is never touched during the preview stage. ⁠APPLY⁠ then uses ⁠commit-tree⁠ + ⁠update-ref⁠ to seal the deal cleanly. Client & Backend Support **13 Clients Configured Automatically:** Runs out of the box with ⁠git-courer mcp setup⁠ for Claude Code, Cursor, Windsurf, OpenCode, Cline, Roo Code, VS Code, Zed, Claude Desktop, Continue, and more. **100% Local-First:** Works with any backend exposing an OpenAI-compatible ⁠/v1⁠ API (Ollama, LM Studio, llama.cpp). The project is fully open source. I’d love to hear your thoughts on the architecture, the plumbing pipeline, or any features you'd like to see added! **Repo:** [github.com/Alejandro-M-P/git-courer](https://github.com/Alejandro-M-P/git-courer)