What is chatgpt_coding?

Local Coding Agent turns a folder on your own computer into a workspace that ChatGPT Web, or any other tool that speaks the MCP protocol, can work inside of directly. Once set up, ChatGPT can read your files, edit code, run commands, check git status and diffs, and report back on the health of the local setup, acting like a coding assistant that operates on your real machine rather than a sandbox. The project ships with a setup control panel you open in a browser to configure everything: which folder to work in, which network ports to use, and how cautious the agent should be. Two safety modes are offered, a safe mode that blocks many risky command patterns, and a balanced mode that allows normal editing and testing but asks for local approval before doing something risky like deleting files or making network installs. The README is explicit that this tool is not an operating system sandbox: it can genuinely read, write, and run commands on your computer, so it should only be pointed at folders you trust. Getting connected involves several pieces: Node.js 18 or newer, a separately downloaded tunnel client from OpenAI that creates a secure connection between your machine and ChatGPT, a runtime API key with permission to use tunnels, and a tunnel ID. After running an install script and placing the tunnel client in the right folder, you start the setup UI, fill in the workspace path and connection details, and save and start the service. From there, you create a custom app inside ChatGPT's settings pointed at your tunnel, after which new chats can call tools exposed by the local agent, such as asking it for information about the connected workspace. The README also documents a command-line path for people who prefer not to use the browser UI, a table of common error messages and how to fix them, and a test suite for checking the agent's behavior, with a warning not to run the security tests against a real project folder since some of them intentionally test deletion and blocking behavior.