Chat On Steroids gives AI chat apps local coding tools over MCP
Chat On Steroids is a tray app and Chrome extension that turns any ChatGPT conversation into a full coding environment, with file edits, shell commands, worker chats and sessions that continue past the context window limit, at 629 GitHub stars in 12 days.
Image: GitHub
Why it mattersA team already paying for ChatGPT gets Codex-style coding tools for the model without switching to a separate agent product, so the question becomes governance rather than tooling choice.
Chat On Steroids is a local MCP server, tray app and Chrome extension that give ChatGPT the tools OpenAI's own Codex CLI uses: read, search, patch files, run shell commands, and control an interactive terminal. The project is at 629 GitHub stars on a repository created on 22 August, with release binaries and SHA256 hashes for Windows, macOS and Linux.
Codex-grade tools through MCP
The tool contracts for apply_patch, exec_command and write_stdin are ported from OpenAI's Codex CLI, so a model already trained on those formats does not have to relearn them for this server. Multi-file patches are checked before anything is written. Commands run as real operating system processes with interactive stdin, output budgets, and background results a later turn can collect.
The Core capability set exposes at most seven tools at once inside ChatGPT, covering read, search, patch, shell, sessions and workers. A separate Desktop connector adds screen capture, mouse, keyboard and clipboard on Windows and macOS, and is a switch a user turns on rather than a default.
Sub-agents and sessions that last longer than one chat
A main ChatGPT chat can spawn worker chats through the agents tool, hand them tasks, read their reports, and resume them later. Workers are ordinary ChatGPT conversations in the same browser, managed by the app, so a user can watch every one.
Every tool call and its real result is recorded locally, with 30 day retention by default. When a chat gets close to its context limit, a Compact and Resume flow asks it for a handover summary, opens a new conversation, and reconnects to the same local session, so the model can still query what the earlier chat did. Goal and Loop modes give the job of typing "keep going" to a second model, so long unattended runs stop needing a person to prompt each turn.
What it needs, and what it costs to adopt
The ChatGPT side needs Developer mode and custom MCP apps enabled, which OpenAI currently documents as a beta for Business, Enterprise and Edu workspaces, with Pro limited to read and fetch. A Business workspace needs an admin to switch it on. The Chrome extension runs only on chatgpt.com, chat.openai.com and the app's own loopback bridge.
The app is a beta with real permissions. exec_command runs programs as the logged-in user, so the README asks a user to approve a project folder rather than a home directory, and to work on code that is committed somewhere. Release binaries are not publisher-signed and macOS builds are not notarised, so Gatekeeper and SmartScreen will warn on first run.
If the team already pays for ChatGPT, this is a quick way to get Codex-like coding tools for that same account, without adding a second agent product to the invoice. It fits teams whose developers already rely on the ChatGPT app for planning and reading, and who want the same conversation to also edit the code.
The tool forces a governance decision. A local MCP server that can patch files, run commands and control the desktop needs the same review a self-hosted agent would get: which folders it may see, whether the Desktop connector is on at all, whether Goal and Loop are allowed for unattended work, and where the local session history lives. The controls exist; the failure mode is skipping them before the first agent run.
Source
Primary source: totec448-spec/chat-on-steroids on GitHub.
This item was written by an AI system from the linked source. Reveneau is responsible for what it publishes.
Get AI News in your inbox
New developer tools, model and agent releases, and how teams are actually using them to release software. Short, and only when there is something worth reading.