CodexPro Releases Local MCP Server
Key point
CodexPro, a local MCP server that exposes only user-approved repositories to ChatGPT, has been released.
Details
CodexPro is a local MCP server for handling local code within the ChatGPT web interface, exposing only repository paths explicitly approved by the user to the model. Access outside allowed root paths, path traversal via symbolic links, and paths matching blocked globs are all denied.
Key Features and Installation
- Installation: Install via
npm install -g codexproin a Node.js 20+ environment, then runcodexpro setup. - Connection: Enter the generated server URL into the ChatGPT plugin settings to connect. Authentication is token-based within the URL; sharing tokens is prohibited.
- Tunneling: Supports Cloudflare, ngrok, and Tailscale tunnels to meet ChatGPT web's HTTPS requirements, and also provides a local-only mode.
- Task Scope: Supports file reading/searching/editing, executing allowed bash commands, reviewing diffs, and writing planning documents.
Safety and Security
- Defaults: A minimum 24-byte token is required when using public tunnels, and bash runs in safe mode by default.
- Blocked Targets: Sensitive paths such as
.env, key files,.git, and build caches are included in the default block list. - Fail-Safe Controls: Provides controls such as failing HTTP requests when tokens are missing, denying path traversal, and forcibly terminating bash child processes.
- Caution: Since package scripts can still be executed even in safe mode, using the
--no-bashoption is recommended for untrusted repositories.
Differences from Similar Projects
Unlike DevSpace or OpenAI Secure MCP Tunnel, CodexPro allows users to declare access scope at the repository root level and is designed so that editing tools are not listed at all unless write mode is enabled. This prevents flows where the model attempts to write and is then denied.
This summary was generated automatically by AI. Check the original for the author's claims and context. Copyright belongs to the original author.
Our guide explains how the AI works. Report summary errors, attribution issues, or removal requests via Contact.