DevInsight MCP

A lightweight Model Context Protocol (MCP) server that gives Claude the ability to inspect, analyze, and grade local Git repositories.

Built with Anthropic's official Python MCP SDK, DevInsight demonstrates all three MCP primitives — Tools, Resources, and Prompts — through practical developer workflows: language stats, TODO tracking, git history, large-file detection, an overall repo health score, and technology-stack fingerprinting.

Demo

DevInsight demo

A live session driving the DevInsight tools against this very repository — repo stats, TODO scan, and large-file detection.

✨ Features

📊 Repository statistics — languages, file counts, line counts
📝 Scan projects for TODO, FIXME, HACK, and XXX
📜 Summarize recent Git commits (author, date, message, +/- lines)
📁 Detect oversized source files that are due for a refactor
🩺 Score overall repository health (README/LICENSE/tests/git, TODO density, file size) with concrete recommendations
🧰 Fingerprint a project's tech stack — languages, frameworks, databases, package managers, CI/CD, deployment
🌳 Browse repositories through an MCP Resource
🤖 Review and prioritize TODOs using an MCP Prompt

Scanning automatically skips noise: .git, .github, node_modules, virtualenvs, build/cache directories, lock files, and binary assets — see Repository Scanning below.

Why DevInsight?

Developers spend a surprising amount of time manually inspecting repositories:

searching for TODOs
checking Git history
counting files
finding oversized modules
judging whether a project is in good shape before diving in

DevInsight exposes these tasks as MCP tools so Claude can perform them for you, directly in conversation.

Instead of manually searching your project, you can simply ask:

"Summarize the last 10 commits."

"Find every TODO and tell me which ones are most important."

"Which files are becoming too large?"

"How healthy is this repo, and what should I fix first?"

"What's the tech stack of this project?"

Installation

Install from PyPI (easiest)

pip install devinsight-mcp
devinsight-mcp        # runs the server over stdio

Or run it without installing, using uv:

uvx devinsight-mcp

From source (development)

Clone the repository

git clone https://github.com/AzalHoseinian/devinsight-mcp.git
cd devinsight-mcp

Create a virtual environment

python3 -m venv .venv
source .venv/bin/activate

Windows:

.venv\Scripts\activate

Install dependencies

pip install -r requirements.txt

Run the tests

pip install pytest
python -m pytest -q

Try it with the MCP Inspector

The easiest way to test the server standalone is with the official MCP Inspector:

mcp dev server.py

The Inspector lets you invoke every Tool, inspect Resources, test Prompts, and debug raw responses — all in the browser, no client app required.

Claude Desktop Setup

Add DevInsight to your claude_desktop_config.json:

macOS: ~/Library/Application Support/Claude/claude_desktop_config.json
Windows: %APPDATA%\Claude\claude_desktop_config.json

Easiest (uses uvx):

{
  "mcpServers": {
    "devinsight": {
      "command": "uvx",
      "args": ["devinsight-mcp"]
    }
  }
}

From a clone:

{
  "mcpServers": {
    "devinsight": {
      "command": "/absolute/path/to/devinsight-mcp/.venv/bin/python3",
      "args": [
        "/absolute/path/to/devinsight-mcp/server.py"
      ]
    }
  }
}

If you installed the command with pip install . or pip install devinsight-mcp, you can point Claude Desktop at it instead:

{
  "mcpServers": {
    "devinsight": {
      "command": "/absolute/path/to/devinsight-mcp/.venv/bin/devinsight-mcp"
    }
  }
}

Using the virtual environment's Python interpreter (rather than a bare python) ensures Claude Desktop finds the mcp package regardless of what's active in your shell.

Fully quit and reopen Claude Desktop afterwards — it only reads this file on startup. Then try:

Use repo_stats on ~/projects/my-app

Review all TODOs in this repository.

Security notes

DevInsight is read-only: it never writes, deletes or executes anything from the repositories it scans. The only external command it runs is git log with a fixed argument list.

It can, however, read any path you give it that your user account can read, so only connect it to clients you trust. It does not follow symlinked directories when drawing the tree, and it skips lock files and binaries in statistics.

Available Tools

Primitive	Name	Purpose
Tool	`repo_stats`	Language and line-count breakdown
Tool	`find_todos`	Find TODO / FIXME / HACK / XXX comments
Tool	`git_log_summary`	Summarize recent Git activity
Tool	`find_large_files`	Detect files exceeding a configurable size
Tool	`repo_health`	Overall 0-100 health score with recommendations
Tool	`tech_stack`	Detect languages, frameworks, databases, package managers, CI/CD, deployment
Resource	`repo://tree/{path}`	Render the repository tree
Prompt	`review_todos`	Ask Claude to prioritize TODOs

Usage examples

repo_stats(path="~/projects/my-app")
→ { "total_files": 142, "total_lines": 18734,
    "by_extension": { ".ts": {...}, ".tsx": {...}, ... } }

find_todos(path=".", max_results=50)
→ [ { "file": "src/api.ts", "line_number": 42,
      "tag": "TODO", "text": "handle retry backoff" }, ... ]

git_log_summary(path=".", count=5)
→ { "commits": [ { "hash": "a1b2c3d4", "author": "...", 
      "date": "2026-07-10", "message": "...", 
      "insertions": 12, "deletions": 3 }, ... ] }

find_large_files(path=".", threshold_lines=300)
→ [ { "file": "src/legacy/parser.py", "lines": 812 }, ... ]

repo_health(path=".")
→ { "score": 78,
    "checks": { "has_git": true, "has_readme": true,
                "has_license": true, "has_tests": false },
    "recommendations": [ "Add a test suite ..." ] }

tech_stack(path=".")
→ { "languages": ["Python"], "frameworks": ["FastAPI"],
    "databases": ["PostgreSQL"], "package_managers": ["pip"],
    "ci_cd": ["GitHub Actions"], "deployment": ["Docker"] }

Repository Scanning

Statistics tools (repo_stats, find_todos, find_large_files, repo_health) walk the repo while pruning directories as they go — ignored subtrees are never descended into.

Ignored directories: .git, .github, node_modules, venv, .venv, env, __pycache__, dist, build, .next, .nuxt, .svelte-kit, .idea, .vscode, target, coverage, htmlcov, .pytest_cache, .mypy_cache, .cache

Excluded from statistics: lock files (package-lock.json, pnpm-lock.yaml, yarn.lock, poetry.lock, Cargo.lock, Gemfile.lock, composer.lock, uv.lock, bun.lockb) and binary assets (.png, .jpg, .jpeg, .gif, .pdf, .zip, .exe, .dll) — these are real project content, but noise in line/size stats.

tech_stack looks at these same files directly, since a lock file's mere presence is a useful package-manager signal.

Architecture

               Claude Desktop
                      │
                      ▼
             DevInsight MCP Server
                      │
      ┌───────────────┼───────────────┐
      ▼               ▼               ▼
 Git Repository   File System     Git History

Project Structure

devinsight-mcp/
├── devinsight_mcp.py      # implementation (installed as the `devinsight-mcp` command)
├── server.py              # thin launcher for `mcp dev server.py`
├── tests/
├── .github/workflows/test.yml
├── pytest.ini
├── requirements.txt
├── pyproject.toml
├── README.md
├── docs/
│   ├── demo.gif
│   ├── demo.tape
│   └── demo_cli.py
└── .gitignore

Roadmap

Tech Stack

Python
Anthropic MCP Python SDK
Git
Claude Desktop
Model Context Protocol (MCP)

Contributing

Contributions, suggestions, and feedback are welcome.

If you'd like to improve DevInsight, feel free to open an issue or submit a pull request.

License

MIT — see LICENSE.

devinsight-mcp

Documentation