RA Browser Use

An AI web agent for Chrome that runs entirely on your machine.

Give it a task in plain language and it navigates, clicks, types, and extracts for you — with the model running on your own hardware. No API key. No cloud. No page content leaving your browser.

Forked from Nanobrowser · Powered by RunAnywhere · Apache-2.0

---

This is a public fork of Nanobrowser

RA Browser Use is an openly-declared fork of Nanobrowser (Apache-2.0), adapted for local, on-device AI and powered by the RunAnywhere SDK.

We want to be unambiguous about that, so:

• Upstream's full commit history is merged into this repository, not squashed away. git log and git blame show the real provenance of every inherited file.

• Upstream is tracked as a live git remote, and we pull its fixes forward:

  git remote add upstream https://github.com/nanobrowser/nanobrowser.git
  git fetch upstream && git merge upstream/master

• Attribution and the list of significant modifications are in NOTICE.

Nanobrowser built an excellent multi-agent browser automation extension. Our contribution is a different answer to where the model runs: Nanobrowser connects to cloud LLM providers (or a separately-installed Ollama), while RA Browser Use's default is a model executing inside the extension itself. Credit for the agent architecture, the DOM serialization, and the extension foundation belongs upstream.

What changes in this fork

	Nanobrowser	RA Browser Use
Where the model runs	Cloud API, or a separate local Ollama process	Inside the extension, via RunAnywhere (WebAssembly + WebGPU)
Setup	Bring your own API key	Nothing to configure — it picks a model for your hardware
Page data	Sent to whichever provider you configure	Never leaves the browser on the default path
Cost	Per-token provider billing	Free to run
Cloud providers	The product	Still supported, now optional

Cloud providers are deliberately kept. Local inference is the default and the point, but nothing stops you pointing RA Browser Use at a frontier model when a task genuinely needs one.

Documentation

• docs/MODELS.md — what ships, why, what was ruled out and the numbers behind it, and how to add a computer-use model.
• docs/TESTING.md — how to build, load and test it, where the logs are, and what each automated gate does and does not prove.

Status

Early development, not yet released. The fork foundation is in place and the extension builds and runs; the on-device inference backend is being wired up. There is no Chrome Web Store listing yet. Track progress in the issues.

Development

Requires Node (see .nvmrc) and pnpm — engine-strict is on, so a mismatched toolchain fails install rather than breaking mysteriously later.

nvm use
pnpm install
pnpm build        # or: pnpm dev  (watch mode)

Then load it in Chrome: chrome://extensions → enable Developer mode → Load unpacked → select dist/.

pnpm type-check                  # tsc across every workspace
pnpm lint
pnpm -F chrome-extension test    # vitest

Chrome and Edge only. Firefox and Safari are not supported — on-device inference here depends on Chrome's offscreen-document, WebGPU, and OPFS behaviour.

Architecture notes for contributors and coding agents live in CLAUDE.md / AGENTS.md.

Privacy

On the default local path, the model runs in the extension, so page content, prompts, and model weights stay on your machine. Weights are downloaded once from a public model host and cached on disk.

Two honest caveats. First, browser automation needs broad browser access, and those Chrome permissions are real regardless of where inference happens. Second, if you configure a cloud provider, page content goes to that provider on the same terms as any other client — that is the trade you are choosing when you do it.

License

Apache-2.0 — see LICENSE, inherited from Nanobrowser and retained. Attribution and modifications: NOTICE.