CodexPro: Turning ChatGPT into a Repo-Scoped Agent

Giving a web-based LLM access to your local filesystem is usually a recipe for anxiety, but CodexPro handles it by treating the repo as a strict sandbox. Instead of just giving ChatGPT a raw shell, it implements a local MCP server that limits the AI to a specific toolset. It's basically a "guardrail-first" approach to an AI workflow.

The Setup Process

Getting this running is straightforward, but the connection logic is where it gets specific. You aren't just running a script; you're exposing a local endpoint to ChatGPT's app settings.

1. Install the CLI globally:

npm install -g codexpro

2. Initialize the workspace:

codexpro setup

3. In the ChatGPT web interface, go to Settings → Apps → Advanced settings → Create app. Paste the Server URL provided by the CLI and set authentication to "None" (the tool uses its own URL token for security).
4. Launch the session:

codexpro start

Why the Tooling Matters

Most "AI agents" just dump a terminal in front of the model. CodexPro is more disciplined. The agent gets a defined API to interact with your code:

Read/Search: Standard file access.

Edits: write, edit, and apply_patch for making changes.

Verification: A bash tool that blocks dangerous patterns and secret paths.

Review: show_changes to verify what actually happened.

The "handoff" system is the real MVP here. If the agent wants to execute something complex, it writes a plan to a .ai-bridge directory. The AI can't just trigger the execution; you have to manually run codexpro execute-handoff locally. This separation of planning and execution is a massive win for developer experience and security.

Tunneling and Connectivity

Since ChatGPT needs a public HTTPS URL to talk to your localhost, CodexPro integrates with Cloudflare tunnels, ngrok, and Tailscale. For anyone who cares about stability, the Tailscale Funnel path is the way to go, though it requires MagicDNS and HTTPS certs to be sorted on your tailnet first.

Final Take

If you're already paying for ChatGPT and want a practical tutorial on how to let it edit your actual code without the "black box" feeling of some IDE plugins, this is a solid choice. It doesn't overpromise—it doesn't claim to be a magic "solve all bugs" button—it just provides a clean, scoped interface.

The project is young, but the focus on memory buffers (like the 64 KiB binary check) and strict path blocking shows it's built by someone who actually codes.

https://github.com/codexpro-ai/codexpro