Open ChatGPT Atlas: a Chromium side panel that drives Gemini Computer Use and Composio tools
Open Source and Free Alternative to ChatGPT Atlas.
At a glance
- What is it?
- Open ChatGPT Atlas is a TypeScript browser extension plus an Electron browser that puts an agent in a side panel, using Gemini 2.5 Computer Use for visual automation and Composio's tool router for app integrations. It installs from source with npm, and it is not a drop-in replacement for OpenAI's Atlas.
- Who is it for?
- Adopt it if you want a readable TypeScript codebase where the agent loop lives in background.ts, content.ts and tools.ts, and you are comfortable supplying your own Google API key. Do not adopt it if you need a packaged, signed desktop browser or a documented rollback path, because the README lists no releases and no recovery procedure.
- Can I use it commercially?
- Not without permission. GitHub finds no licence file in the repository, and without a licence all rights are reserved by default: you may read the code but not reuse it. Check the README, or ask the authors, before using it.
- Is it still maintained?
- Yes. The repository last received commits 56 days ago.
- What is it written in?
- Mainly TypeScript, according to GitHub's language statistics.
Answers come from the project's GitHub data, last synced on September 15, 2026, and from our analysis. They are not legal advice.
DEEP OPEN-SOURCE ANALYSIS
What Open ChatGPT Atlas actually solves, and for whom
OpenAI's Atlas is a browser with an agent built into it. Open ChatGPT Atlas takes the other route: it is an extension you load into Chrome or Edge, plus a separate Electron application, and it recreates the side panel chat experience on top of models you supply keys for. The README describes it as an open source and free alternative to ChatGPT Atlas, and the repository is written in TypeScript.
The intended user is an engineer who wants to see and change the agent loop rather than consume a closed product. Everything in the repository is plain TypeScript: background.ts, content.ts, tools.ts, sidepanel.tsx and settings.tsx, with Vite as the bundler and React for the interface. If you have ever wanted to know how a browser agent decides to click something, this codebase is small enough to read end to end in an afternoon. That is the real value proposition, not feature parity with a commercial browser.
It is a poor fit for anyone who wants a finished desktop product. The README lists no releases, so there is no packaged artifact to download, and the repository does not state a licence, which matters if you plan to ship a derivative.
Two modes, two different automation mechanisms
The feature list separates two things that are easy to confuse. Tool Router Mode uses Composio's routing to reach Gmail, Slack, GitHub and, according to the README, more than 500 integrations. Browser Tools Mode uses Gemini 2.5 Computer Use for visual automation: screenshots, clicks, typing, scrolling and navigation. They are toggled from the same ◉ button in the chat header, and the README notes that turning Browser Tools off lets Composio tools take over.
The architectural consequence is that the two modes fail differently. Tool Router calls are API calls to integrations, so a failure looks like an authentication or scope error. Browser Tools failures are visual: the model misreads a page, clicks the wrong element, or a site's layout defeats the screenshot. The README mentions visual feedback in the form of blue click indicators and element highlighting during automation, which is the practical way to tell which of those two things just happened.
One design point worth noting: the README states that no backend is required and that all API calls are made directly from the extension. That keeps deployment trivial, and it also means your Google and Composio keys live in the extension's own settings storage. The repository ships a .env.example with OPENAI_API_KEY and ANTHROPIC_API_KEY entries, and package.json pulls in @ai-sdk/anthropic, @ai-sdk/google, @ai-sdk/openai and @modelcontextprotocol/sdk. The README's setup instructions, however, only walk through the Google and Composio keys, so the environment file and the documented configuration do not fully line up.
Installing the extension and running a first browser task
The README's installation path is a source build. You need Node.js 18 or newer, npm, and Chrome or Edge with Manifest V3 support. Clone the repository, then install dependencies and build. The build output goes to dist, which is the folder you load into the browser.
npm install
npm run buildAfter the build finishes, open chrome://extensions/, enable Developer mode, click Load unpacked, and select the dist folder. The README then says to open Settings through the gear icon to configure API keys. A Google API key is required; get it from Google AI Studio and paste it under Google API Key. A Composio API key is optional and only needed for Tool Router mode.
For a first real use, click the ◉ button in the chat header to enable Browser Tools, then give a natural language instruction. The README's own examples are the safest starting point.
Navigate to reddit.com and scroll down
Click on the search box and type 'puppies'
Take a screenshot of this pageWhat you should see is the page reacting under the agent's control, with the click indicators described in the feature list marking where the model acted. If nothing happens, the README points to FAQ.md and TROUBLESHOOTING.md rather than to the main document, which is a hint about where the common failures live.
If you want the standalone browser instead of the extension, the README gives a separate path. Note that the scripts block in package.json lists dev, build and preview only.
npm run build:electron
npm run electronThe README also documents npm run electron:dev for hot reload and npm run dev for extension development with a manual extension reload after each change.
Where this breaks down
The honest limitation is that the README never documents rollback, and the repository lists no releases. If an agent action goes wrong, there is no described undo. The safety feature listed is confirmation dialogs for sensitive actions such as checkout and payment, which is a guard before the fact, not a recovery after it. For anything involving a logged-in account, that distinction matters.
The second limitation is key handling. Because calls go directly from the extension, your Google key sits in browser-local settings. Anyone with access to that browser profile can read it. There is no described proxy or server-side key exchange, and adding one would contradict the no-backend design.
The third is model dependency. Browser Tools is tied to Gemini 2.5 Computer Use Preview, and the README describes it as a preview model. Preview models change. The package.json lists @ai-sdk/anthropic and @ai-sdk/openai alongside @ai-sdk/google, so the codebase has more provider surface than the documentation explains, but the documented browser automation path is Gemini only.
Finally, this is the wrong tool if you need a hardened, signed browser. It is an unpacked extension loaded in developer mode, and the Electron variant is built and launched from source.
How it compares to Browser Use and Playwright-style automation
The closest conceptual alternative is Browser Use, which drives a browser through a Python agent loop and is typically run as a script or service you control. The difference in approach is where the agent lives. In Open ChatGPT Atlas the agent lives inside the browser as an extension, sharing the user's session and cookies, and the interaction surface is a side panel you talk to. In Browser Use the agent is external and the browser is something it controls.
That changes the failure profile. An in-browser extension inherits your logged-in state, which is convenient for tasks like reading Gmail and risky for the same reason. An external agent usually starts from a fresh profile and has to authenticate explicitly.
The other comparison is Playwright. Playwright is deterministic: you write selectors and assertions, and it does exactly that. Open ChatGPT Atlas is probabilistic: the README's examples are natural language instructions, and the model decides which element to click. For a regression test suite, Playwright is the right answer and this is not. For a task you cannot write selectors for because the page changes, the model-driven approach is the only one that works at all.
Maintenance, upgrades and licensing
The repository is not archived, and the last push was on 2026-07-21. That is a recent commit history, so the code is being touched, but a single push date does not tell you the release cadence. With no published releases, upgrades are done by pulling the branch and rebuilding, not by installing a versioned artifact. Budget for that: every upgrade is npm install, npm run build, then reload the unpacked extension in chrome://extensions/.
Dependency churn is the real upgrade cost. The package.json pins major versions of the AI SDK packages, @ai-sdk/google at ^3.0.0 and ai at ^6.0.0, plus @modelcontextprotocol/sdk at ^1.0.4. Caret ranges mean a fresh npm install can pull newer minor and patch versions than the author last built against, and the Gemini Computer Use integration is the part most exposed to that drift.
The repository does not state a licence. That is not a detail you can defer: without a stated licence, the default position is that no rights are granted, and the README's own framing as a free alternative does not change that. Check the repository's licence file before you fork, redistribute or ship anything built on it. This is a factual gap, not legal advice.
Editorial conclusion
Adopt it if you want a readable TypeScript codebase where the agent loop lives in background.ts, content.ts and tools.ts, and you are comfortable supplying your own Google API key. Do not adopt it if you need a packaged, signed desktop browser or a documented rollback path, because the README lists no releases and no recovery procedure. Before installing, open package.json and confirm the scripts block: it defines dev, build and preview, while the README also calls npm run build:electron and npm run electron, which are worth checking against the repository state you clone.
Frequently asked questions
How do I open Open ChatGPT Atlas after installing it?
It is not a standalone app in the extension path. After npm run build, you load the dist folder through chrome://extensions/ with Developer mode on, and the interface appears as a side panel chat you open from the browser.
Why can't I open Open ChatGPT Atlas?
The most likely cause is that the extension was never built or the wrong folder was selected. The README specifies selecting the dist folder, not the repository root, and it points to TROUBLESHOOTING.md for common issues.
Is Open ChatGPT Atlas free, and what does it cost to run?
The repository describes it as open source and free. You still need a Google API key for Gemini, which is required, and an optional Composio API key for Tool Router mode, so model and integration usage is billed by those providers.
Does Open ChatGPT Atlas need a backend server?
No. The README states that all API calls are made directly from the extension and that no backend is required, which is why your API keys are entered in the extension's own Settings panel.
What is the difference between Browser Tools and Tool Router mode?
Browser Tools uses Gemini 2.5 Computer Use to control the page visually with screenshots, clicks, typing and scrolling. Tool Router uses Composio to call integrations such as Gmail, Slack and GitHub, and the README says it covers more than 500 integrations.
Can I run Open ChatGPT Atlas as a desktop browser?
Yes, the repository includes an Electron browser under electron-browser/. The README gives npm run build:electron followed by npm run electron, and npm run electron:dev for development with hot reload.
Community notes