DevPilot MCP

Chat with a code repository and watch exactly what the AI does. DevPilot lets an AI model read, search and inspect one repository through 18 safe tools, and it never changes a file or runs tests without your approval.

It has two parts:

• DevPilot Studio: a local web UI. This is the easiest way to use it (see below).
• DevPilot MCP server: the same 18 tools for any MCP client, such as Claude Code or Gemini.

Watch the full 85-second demo (MP4)

Quick start (recommended: the web UI)

Requires Python 3.10+ (3.11+ recommended). Git is needed for the Git tools.

pip install "devpilot-mcp[ui]"
$env:DEVPILOT_WORKSPACE = "C:\path\to\your\repo"     # macOS/Linux: export DEVPILOT_WORKSPACE=/path/to/your/repo
devpilot-ui

Your browser opens at http://127.0.0.1:8765. Then:

1. Try it right away, no account needed. Click a quick action (Analyze, Git status, Commits), press Ctrl+K to run any tool, or open the Files tab to browse the code.
2. Turn on AI chat (optional). Get a free token at aipipe.org/login, then set it and restart:

   $env:AIPIPE_TOKEN = "your-token"     # macOS/Linux: export AIPIPE_TOKEN=your-token
   devpilot-ui
   

3. Ask a question, like "Where is authentication implemented?" Each answer shows the steps the AI took.

Without a token, everything works except the chat. The full guide is in the Web UI docs.

What DevPilot Studio gives you

• Steps you can follow: every answer lists the tools the AI used, with timings. Anything DevPilot's security refuses (for example reading .env) is shown as blocked.
• Result cards, not JSON: language bars, commit timelines, colored diffs, grouped search matches, test results.
• Ctrl+K tool runner: run any of the 18 tools through a form, no AI needed.
• File explorer: browse the repository, open files, and ask the AI about one.
• Changes with Undo: every patch the AI applied is listed with its diff and a one-click undo.
• Approval first: changes and test runs always wait for you to click Approve.

Steps you can follow	Result cards, secrets blocked	Approval with the exact diff

Use DevPilot from an AI client (MCP)

The same tools work in any app that supports MCP over stdio. With Claude Code, for example:

pip install devpilot-mcp
claude mcp add devpilot -e DEVPILOT_WORKSPACE=C:/path/to/your/repo -- devpilot-mcp

Then ask: "Use DevPilot to analyze this repository and explain its structure." You can also explore the tools by hand with MCP Inspector (npx @modelcontextprotocol/inspector). More in the usage guide.

Always set DEVPILOT_WORKSPACE to the absolute path of the repository. If devpilot-mcp is not on your PATH (for example it is inside a virtual environment), give claude mcp add the full path to the devpilot-mcp executable instead.

The 18 tools

DevPilot collects evidence and performs tightly bounded actions. The AI model does the reasoning, and it never writes a patch on DevPilot's behalf: patches come from the model or from you.

Security at a glance

• Every path stays inside the workspace. .., absolute, drive and UNC paths, and escaping symlinks and junctions are rejected.
• .env files, private keys, credential files and .git internals are never read, searched or written.
• Changes and test runs need your explicit approval in the UI, and patches are validated in full and applied atomically.
• Only two places start processes: allowlisted read-only Git commands, and two fixed Python test commands. There is no shell and no caller-supplied command.
• GitHub access is HTTPS GET to api.github.com only, and the token never appears in output.
• The UI listens on 127.0.0.1 only, and your AI token stays on the server side.

It is not a sandbox. Tests run with DevPilot's own permissions and network access, and secret redaction is best-effort. See Security for the full model and known limitations.

Run from source

git clone https://github.com/KD-kaustubh/devpilot-mcp.git
cd devpilot-mcp
python -m venv .venv
.venv\Scripts\activate           # macOS/Linux: source .venv/bin/activate
pip install -e ".[ui]"
python -m unittest discover -s tests -t .

The UI frontend needs Node.js to rebuild (cd ui && npm install && npm run build); see Development.

Documentation

License

MIT