ShadowGit MCP: Automate Git History for Efficient AI Coding

If you use AI coding assistants like Claude, Cursor, or Copilot, giving them the right context is key.

ShadowGit captures every save as a minute-by-minute git commit in a hidden repo and exposes that history to AIs via an MCP server, reducing token usage by about 66% by letting the model query specific changes instead of re-reading entire files.

Think of it like a time-lapse camera for your code—fast to review and private to your machine.

Why ShadowGit + MCP matters

Token reduction: AI assistants query precise diffs and commits instead of re-parsing large codebases; ShadowGit reports an estimated ~66% token reduction by design.
Automated backups: Every save becomes a commit; no more manual commit anxiety.
Local-first security: Your history stays on your machine—no cloud uploads by default.
MCP compatibility: Uses the Model Context Protocol (MCP) so multiple assistants can access the same history interface.

Who benefits most

This tool is aimed at developers and teams who use AI-assisted development and want accurate historical context without extra commit overhead.

Individual developers using Claude or Copilot for debugging and code review
Startups accelerating dev cycles with AI-driven PRs and fixes
Security-conscious teams that require local-only code history

How ShadowGit works (quick overview)

ShadowGit runs on your machine, captures saves as commits into a hidden repository (usually .shadowgit.git), and runs a small MCP server that exposes read-only git operations to authorized local AI clients.

The MCP server restricts to safe git queries (logs, diffs, blame, show) and never pushes your history to remote clouds.

Key components

Watcher: Captures file saves and creates commits.
Hidden Git repo: Stores the minute-by-minute history without interfering with your main .git.
MCP server: A local process that translates MCP requests into safe git commands and returns structured results to the AI.

Installation and setup

Below is a step-by-step setup you can follow. If you hit snags, the project repo has issues and a community that can help: ShadowGit MCP on GitHub and the product site: shadowgit.com.

Prerequisites

Node.js (LTS). Verify with node --version.
Claude Desktop (for Claude users) or other MCP-capable AI clients.
ShadowGit app installed and configured to track repos.

MCP server setup for Claude Code

In Claude Code CLI you can register a local MCP server:

claude mcp add shadowgit -- shadowgit-mcp-server
# Restart Claude Code to load the server

Claude Desktop configuration

Edit your Claude Desktop config file and add the server entry. Locations:

macOS/Linux: ~/.config/Claude/claude_desktop_config.json
Windows: %APPDATA%\Claude\claude_desktop_config.json

{
  "mcpServers": {
    "shadowgit": {
      "command": "shadowgit-mcp-server"
    }
  }
}

Install the MCP server from source

If you prefer building from source, the GitHub repo contains the code and tests.

git clone https://github.com/blade47/shadowgit-mcp-server.git
cd shadowgit-mcp-server
npm install
npm run build
npm test

See the upstream project for latest instructions: ShadowGit MCP.

Common commands & usage examples

ShadowGit exposes a small set of read-only git-like operations through the MCP interface.

// View recent commits
await shadowgit.git({
  repo: "my-project",
  command: "log --oneline -10"
})

// Check what changed recently
await shadowgit.git({
  repo: "my-project",
  command: "diff HEAD~5 HEAD --stat"
})

// Find who changed a specific line
await shadowgit.git({
  repo: "my-project",
  command: "blame src/auth.ts"
})

Available tool actions

Repository listing: Get all ShadowGit-tracked repos.
Safe git commands: Logs, diffs, blame, show, and other read-only queries.

Deep dive: How using history reduces token usage

Large language models spend tokens encoding context. When an assistant reads a full file, that consumes tokens each time.

ShadowGit lets the model request just the relevant commit, diff, or blame output so the assistant loads small snippets instead of entire files.

Empirically this reduces token consumption by about 66% for typical workflows that rely on recent edits.

Security and privacy considerations

ShadowGit was designed with local-first privacy. Your hidden repo and MCP server run on-device; nothing is uploaded by default.

MCP servers are configured to only execute safe, read-only git commands. If you need enterprise controls, run ShadowGit behind local policies or on isolated developer workstations.

For more on the protocol and security model, see the MCP quickstart.

Real-world use cases and workflows

Use case: Faster code reviews

Instead of pasting file contents into the assistant, ask it to review a specific commit range. The AI fetches only those diffs, explains intent, and suggests changes while using far fewer tokens.

Use case: Debugging with history

When a bug appears, request the commit that introduced a failing line and the blame. The assistant can focus on that change and propose fixes without re-reading the entire module.

Use case: PR contextualization

Draft PR descriptions by summarizing a set of automatic commits—ShadowGit's granular history gives the AI exact context to produce accurate summaries.

Comparison: ShadowGit vs. conventional workflows

Feature	Traditional git	ShadowGit + MCP
Commit frequency	Manual, variable	Automatic, minute-by-minute
AI context cost	High — model re-reads files	Low — model queries diffs/commits
Privacy	Depends on push targets	Local-only by default

Troubleshooting & tips

Spotted something odd? Open an issue on the project repo and include a tiny repro and the failing commit hash.
If the MCP server doesn’t appear in Claude, restart the app after adding the server entry.
Limit MCP access to trusted local processes—treat the MCP socket like any other dev-secret.

FAQ

Will this interfere with my existing git repo?

No. ShadowGit stores commits in a hidden repo (.shadowgit.git) and operates independently of your working .git.

Does ShadowGit upload code anywhere?

No—by default it keeps history locally. Any networked sharing would require explicit configuration.

Can other AI assistants use the same history?

Yes. Because ShadowGit exposes history via MCP, any MCP-capable assistant can request the same read-only data model.

Next steps and recommended experiments

You don’t need to overhaul your workflow overnight. Try this small experiment: enable ShadowGit on one repo and compare assistant costs for a week.

Notice how many fewer tokens are used when the AI retrieves specific commits vs re-parsing files. Try to fix the same bug with fewer prompts using commit-level context.

Resources

Wrapping up

ShadowGit + MCP gives AI assistants precise, private access to your development history, reducing token usage, improving accuracy, and keeping your code local.

Try it on a single project first and measure token and time savings. Start with the repository on GitHub and the MCP docs at modelcontextprotocol.io.

Quick tip: treat ShadowGit like a time-lapse of your work. It helps the assistant answer better questions faster.

Keep building, and don’t hesitate to share feedback—tools like this improve fastest when we try them in the wild.

ShadowGit MCP: Automate Git History for Efficient AI Coding