Chat On Steroids Turns ChatGPT Into a Full Coding Agent on Windows

A local MCP bridge turns ChatGPT into a Windows coding agent with file access, shell commands, desktop control, and multi-agent worker tabs.

AlphaSignalAlphaSignal
·
·
Chat On Steroids Turns ChatGPT Into a Full Coding Agent on WindowsPRO
AuthorBy AlphaSignal Newsroom
Read1 min
TypeRepo
TopicAgents · Development
SubtopicComputer Use · Mcp · Tool Use
  • Chat On Steroids is a Windows app that bridges ChatGPT to local files, shell, and desktop via MCP.
  • Two connectors expose read, apply_patch, exec_command, screen capture, mouse/keyboard, and clipboard tools.
  • Compact and Resume auto-hands off long chats to a new conversation at ~400k tokens, preserving local session state.
  • Optional goal loop uses an OpenRouter model to auto-continue ChatGPT until a stated task is complete.
  • Experimental multi-agent mode spawns up to eight worker ChatGPT tabs coordinated by a prime chat.
  • MIT-licensed, unsigned installers; full MCP write support requires ChatGPT Business, Enterprise, or Edu.

A new open-source project called Chat On Steroids aims to close the gap between ChatGPT's browser UI and the local, tool-driven workflows offered by Codex, Claude Code, or Cursor. It ships as a Windows desktop app paired with a Chrome extension, using OpenAI's Model Context Protocol (MCP) to give ChatGPT scoped access to your filesystem, shell, and desktop.

The design keeps ChatGPT running in the browser while a local app acts as the permission boundary and executor for anything the model wants to touch on your machine. The repo is MIT-licensed and distributed as prebuilt installers.

What the bridge exposes

The app publishes two MCP connectors that ChatGPT talks to over a tunnel. The Core connector covers approved files, search, patches, terminal, session lookup, and workers, through tools named read, view_image, find, apply_patch, exec_command, write_stdin, session, and agents. The Desktop connector covers screen, windows, mouse/keyboard, and clipboard via observe and computer.

The security model is defense-in-depth rather than a hard sandbox. Key boundaries to understand:

  • File tools are limited to approved folders, with paths validated and canonicalized before access, but containment is application-level, not a VM.

Pro article

This story is for Pro members

You've reached the end of the free preview. Upgrade to AlphaSignal Pro to read the full article - and everything else behind the paywall.

Comments

avatar