Chat-on-Steroids Gives ChatGPT Real Hands on Your Computer, and a Real Policy Risk

On September 7, a desktop app called chat-on-steroids picked up 860 GitHub stars in a single day, a 130 percent jump that made it the single biggest riser on GitHub's daily radar. The timing lines up with a complaint AI researcher Ethan Mollick posted the same week: it's "getting very hard to track where your work and conversations are" across every device and app an AI assistant touches. Chat-on-steroids is built around fixing exactly that problem for ChatGPT users, and it does it by giving ChatGPT something it was never supposed to have: hands on your actual computer.

What chat-on-steroids actually connects

The app pairs a desktop program with a Chrome extension. The extension watches your browser-based ChatGPT tab and proves to a small local server, running on your own machine, that a given conversation is really yours. Once that link is live, the desktop app opens up a set of tools ChatGPT wouldn't otherwise have: reading and searching files in folders you approve, applying code patches with a preview step first, running shell commands (with a chance to type in more input if the command asks for it), viewing images, and pulling up your own session history to ask about work from days or weeks ago. Everything stays local. The project's own documentation is specific about that boundary: file tools are restricted to folders you've explicitly approved, and a read-only mode exists that turns off writes and command execution with a single switch.

Compact and resume, built for Mollick's exact complaint

The most relevant feature here is one called Compact & Resume. ChatGPT conversations run into a context limit, the same one every long-running chat eventually hits, and the usual fix is to start over and re-explain everything you were doing. Chat-on-steroids automates the handoff instead: it watches your token usage locally, and when a conversation is about to run out of room, it asks the current chat to write up a brief on where things stand, saves that brief, opens a fresh conversation, and carries the same local session and files into it automatically. The project's own documentation notes that both halves of that handoff are tied to checkpoints in the chat itself, specifically so a page refresh can't accidentally submit the same handoff twice.

The app also supports spawning up to eight worker chats from one main conversation, so a task can be split across several ChatGPT threads working in parallel, each keeping its own history. Worker chats can't talk to each other directly, which limits how much real coordination is possible, but for splitting independent chunks of a task it removes a step that used to mean copying context by hand between browser tabs.

The catch, stated as plainly as the project states it

None of this runs through a documented, sanctioned ChatGPT automation API, because no such API exists for what this tool does. The extension works by observing ChatGPT's own web interface, the same interface a person would use by clicking around, and automating what a person would otherwise do by hand. The project doesn't hide this: its own documentation states outright that OpenAI's terms and policies on automated access, rate limits, and permitted use may apply to an account using it, and tells users to read the agreement governing their own account before turning on the extension or multi-agent mode.

That warning isn't hypothetical. Google suspended paying Antigravity subscribers earlier this month for running a comparable third-party automation tool against its own product. There's no announcement that OpenAI has done the same thing to ChatGPT accounts using tools like this one, but the mechanism, and the appetite, both clearly exist. Anyone wiring this into a workflow they depend on is accepting that their account, not just the tool, is the thing at risk if that changes.

There's a second limitation worth knowing before installing anything: commands run with your full user privileges. The app isn't sandboxed, and desktop control isn't limited to a single folder the way file access is. That's a meaningfully different risk profile than a tool that only reads text. On Linux specifically, the desktop-control backend isn't available at all, and the app requires a working Secret Service keyring rather than falling back to an unencrypted store for secrets, which is a deliberate choice but one more thing to check before assuming it'll just work on your setup.

Who this is actually for

Chat-on-steroids is built for someone who already lives inside ChatGPT for real work, hits the context wall regularly, and finds the current fix (start a new chat, re-explain everything) genuinely costly. If that isn't a daily problem for you, the policy exposure and the unsandboxed command execution are a lot of risk to take on for a convenience feature. If it is a daily problem, this is a working, free, open-source answer to it, built by people who are upfront about exactly what they're asking you to accept in exchange.

Sourced from the chat-on-steroids GitHub repository and Ethan Mollick's post on X.