I built an open-source MCP server that gives AI agents access to ALL your local Git repos at once.

Sandeep Suddala — Tue, 23 Jun 2026 13:33:00 +0000

I asked my AI agent "which microservice owns the payment logic?"

It had no idea. I was pointing it at the wrong repo.

So I fixed that by building repobridge - an MCP server that gives any AI agent access to ALL your local Git repos at once.

The problem
I work across dozens of Java microservices every day. Every time I asked Claude Code or GitHub Copilot to help me, I had to manually navigate to the right repo first. Questions like "where is this config key set?" or "which service calls this endpoint?" took 10 minutes of grepping. I wanted to just ask and get an answer.

What repobridge does
Search code across all your repos in one query
Read files, check git status / diff / log - without leaving your AI conversation
Find which repo contains a specific class, method, or config key
Trigger builds (Gradle / Maven / npm) directly from chat
Works with Claude Code, GitHub Copilot, Gemini, Cursor, Windsurf - anything MCP-compatible

What I learned about MCP
The Model Context Protocol is Anthropic's open standard for connecting AI agents to external tools. Think of it like a USB-C port - one protocol, any device.

Building with the Python MCP SDK (specifically FastMCP) was surprisingly simple. You define a tool with a decorator and a docstring, and the protocol handles everything else:

@mcp.tool()
def search_code(repo_name: str, pattern: str, file_glob: str = "**/*") -> str:
    """Search for a regex pattern across files in a repository."""
    # ... implementation

That's it. The SDK generates the JSON schema, handles transport, and exposes your tool to any connected AI agent.

What I learned about Python (as a Java developer)
Coming from Spring Boot, Python felt both liberating and disorienting:

FastMCP is the equivalent of Spring's @RestController - but 10x less ceremony
pathlib.Path is genuinely better than Java's File
Threading primitives look familiar but the GIL changes everything
pyproject.toml is the pom.xml I always wished I had

The hardest part was unlearning verbosity. Python rewards conciseness in a way Java doesn't.

The result
A single-file server with thread-safe caching, optional auth tokens, ripgrep support for fast search, and multi-transport support (stdio for local, SSE / HTTP for remote agents).

It took about a week to build - mostly because I was learning Python and the MCP SDK at the same time.

Open source on GitHub: https://github.com/sundeepsuddala/repobridge

If you're a Java / Spring developer curious about MCP or Python, a small MCP server is the best learning project I've found. The feedback loop is immediate - your AI agent tells you when something is broken.

DEV Community: Sandeep Suddala

I built an open-source MCP server that gives AI agents access to ALL your local Git repos at once.