Home

Codex Free

Let ChatGPT Web Pro work on the code on your own machine — safely, over an outbound-only tunnel, with no public URL.

Codex Free is a local MCP bridge server. It runs on your computer, exposes a curated set of tools (read/write files, run shell commands, git, search, patching, long-running processes, and more), and connects those tools to ChatGPT through OpenAI's native Secure MCP Tunnel. ChatGPT can then act on your project the way Codex does — patch files in place, drive interactive processes, keep a plan across a task — while every file access stays pinned to the directory you chose.

This wiki is for people who run the server — no Rust knowledge required. It takes you from installation, through every command-line argument, to exactly what happens on each tool call.

---

Start here

If you want to…	Read
Get running in five minutes	Quick Start
Install the binary	Installation
Understand every flag and command	CLI Reference
Understand every config option	Configuration
Know what ChatGPT can actually do	Tools Reference
Understand what happens end-to-end	How It Works
Connect it to ChatGPT	Connecting to ChatGPT

Going deeper

• Multi-Project Mode — serve many repositories from one running server.
• Worktree Isolation — let two chats edit the same repo without colliding.
• Review Checkpoints — show ChatGPT only what changed since the last review.
• Context and Memory — keep a plan and notes across chats and restarts.
• AGENTS and Skills — teach ChatGPT your project's conventions and recurring tasks.
• Bridging MCP Servers — re-expose your other MCP servers through this one.
• Security Model — what a tool call can reach, and how to bound it.
• Audit Logging — record who called what, without leaking secrets.
• Troubleshooting and FAQ.

---

What you get

• Structured tools that behave the same on Windows, macOS, and Linux — read_file, write_file, glob, grep, tree, git operations — cheaper and safer than shelling out.
• Codex's own agent tools, faithfully ported — apply_patch, exec_command/write_stdin, view_image, update_plan, clock_*, skills_*.
• Codex's behavioural brief so the client edits carefully instead of rewriting files or running git reset --hard.
• Bounded output so a single tool call never floods ChatGPT's context.
• Persistent plan and notes that survive a new chat or a server restart.
• Native outbound tunnel — the server listens only on 127.0.0.1, protects itself with a random per-process token, and never opens an inbound port.

One-line mental model

You point Codex Free at a directory. ChatGPT gets a set of tools locked to that directory. Every tool call runs on your machine, as your user, inside the boundary you set — and you can watch, bound, and audit all of it.

Read Security Model before pointing it at anything sensitive.

---

Codex Free is a Rust rewrite of the original Bun + TypeScript codex-free, built on tokio + axum and the official rmcp SDK. It's open source under the MIT license.