WebBrain: Open-Source AI Agent for Chrome and Firefox
The extension reads pages through the accessibility tree instead of selectors, runs on local or cloud models, and switched to GPL-3.0 at version 33.0.0.
Illustrative image: a laptop on a desk at dusk, with fine lines of light branching out of the screen like a tree diagram.
WebBrain is an open-source browser extension that reads a page through its accessibility tree and then clicks, types and runs multi-step tasks on request, using either a locally hosted model or a cloud provider.
At a glance
- Ships for Chrome, Firefox and Edge; the Chrome build uses Manifest V3, the Firefox build Manifest V2.
- Three modes per project docs: Ask (read only), Act (click and type), Dev (adds page source, console and network access).
- License per the repository: GPL-3.0-or-later from version 33.0.0; earlier releases were MIT.
- Model choice is open: OpenAI-compatible local servers such as Ollama, llama.cpp, vLLM or LM Studio, or cloud providers.
- At the time of retrieval the project page listed 989 stars and 119 forks — not independently verified by us.
Browser agents tend to break in the same place: they drive a page through CSS selectors, and the next redesign leaves them grabbing at nothing. WebBrain, an open-source extension for Chrome and Firefox, takes a different route according to its documentation and reads the accessibility tree — the structured description browsers already expose for screen readers.
Three modes, not one blank check
The extension splits its permissions into tiers. In Ask mode the agent may only read the page and answer questions. Act allows clicking, typing, navigation and form filling. Dev additionally opens page source, styles, console and network traffic.
That split matters more than it looks. An agent allowed to submit forms sits in a different risk class than one that only summarizes, and the mode is what decides how much a misread instruction can cost.
Bring your own model
WebBrain does not lock users to a single vendor. Per the repository, any OpenAI-compatible server can be attached — llama.cpp, Ollama, vLLM, LM Studio and LocalAI are named. For hosted models the project says the Chromium build carries 106 built-in provider cards. The default is an in-house managed service called Compass 1.0.
Privacy therefore becomes a configuration question rather than a product claim: with a local model the page content stays on the machine, with a cloud key it does not. The project documents that data flow in dedicated privacy and security-model sections.
Firefox is the weaker build
The project is unusually direct about its own gap: the Firefox build is described as meaningfully weaker than the Chrome one because Firefox offers no equivalent of the DevTools Protocol. The documented consequences are shadow-DOM piercing and mouse-event simulation. Nearly all agent code is shared, with the builds diverging mainly at the browser API layer.
Beyond the extension itself, the monorepo contains an MCP server that lets other agents — Claude Code is the example given — hand browser tasks over, plus a standalone LM Studio plugin. Features include saved workflows, per-tab conversation history and a /watch command that polls a page every 60 seconds by default, adjustable between 30 and 120.
What we could not confirm
Only the project's own presentation on its code-hosting platform was available for this report. Every figure cited — 989 stars, 119 forks, version 33.0.0, 106 provider cards — is a project statement captured at the moment of retrieval and has not been confirmed by a second, independent newsroom.
How well the agent performs in daily use also remains open: we had no independent benchmarks, no extension-store install counts and no external security audit. The move from MIT to GPL-3.0-or-later is recorded in the repository, but the available source does not explain the reasoning behind it.
FAQ
What is WebBrain?
An open-source browser extension for Chrome, Firefox and Edge that reads web pages with a language model and, on instruction, clicks, types and works through multi-step tasks.
Does WebBrain need an API key?
Not necessarily. The project ships with a managed service that requires no setup; alternatively you can point it at a local server such as Ollama or LM Studio, or supply your own cloud provider key.
What license is WebBrain released under?
Per the repository, GPL-3.0-or-later from version 33.0.0 onward. Earlier releases were published under the MIT license.
