Executable programs and extensions
Last updated
Unsloth Browser Use: hand a browser to your local model
A tool-calling model opens pages and fills forms
Browser Use hands one browser tab to a local model. The model needs tool calling, and a vision model can also read screenshots. Payments and sign-ins do not proceed without approval, and localhost or file:// addresses do not open.
A local model holding the browser
Browser Use opens a browser pane beside the chat and lets the model move through pages and press buttons inside it. Any model that supports tool calling works locally.
A vision-capable model also reads screenshots; without vision it cannot see images or charts that the page text does not carry.
You can attach a cloud model too, but the reason to use a local one is that page content stays on the machine.
Which model to attach
Browser Use does not pin a model. It needs tool calling and enough memory to load, but in practice the model that calls tools reliably wins.
The documentation uses Qwen3.8 27B, but an 8 GB machine of memory should start with a 2B or 4B tool-calling model, and pick a vision model if it has to read images on screen.
Four-bit is enough. Browser work is many short calls rather than one long context, so a faster file beats a larger context window1.
What works and what does not
Reading pages, following links, clicking labelled buttons and filling forms work with a text model; images and charts need vision.
Passwords and payment details cannot be filled by the model, and uploads are blocked, so sensitive input stays with the person.
| Workload | Text model | Vision model |
|---|---|---|
| Read pages and follow links | Works | Works |
| Click a labelled button | Works | Works |
| Inspect images and charts | No | Works via screenshot |
| Enter passwords or payment details | Person only | Person only |
| Upload a file | No | No |
Install it and switch it on
Install the desktop app with brew on a Mac or winget on Windows; Linux and WSL use the install script.
brew install --cask unsloth
winget install id UnslothAI.Unsloth
curl -fsSL https://unsloth.ai/install.sh | shPick a tool-calling model in the model hub and download a quantization2 that fits the machine, then turn on Browser use from the menu beside the chat input.
Set permissions first
Permissions live under Settings and Browser; start with the narrow default and open exceptions only for sites you use often.
Full access skips several approval prompts3. Blocked sites and downloads stay blocked and scripts still ask, but fewer checks means you should enable it only for the task at hand.
| Setting | What it controls | Values you can pick |
|---|---|---|
| Browsing | Which sites the model may open | Allow, ask, block |
| Actions | Approval before sign-in, sending or buying | Always ask, full access |
| Downloads | Allow, ask or block | Allow, ask, block |
| Unlabelled clicks | Clicks on areas without a named control | Block, allow |
| History | Whether the model reads your history | Ask first, always allow, never |
| Scripts | Approved each time | Ask each time, never run |
Where it gets stuck
The model stays inside its own tab and cannot drive tabs open in Chrome or Safari, so paste the URL into the chat and let it reopen there.
Localhost, private-network and file:// addresses are not valid targets; only public HTTP and HTTPS pages open.
| Symptom | Cause | Fix |
|---|---|---|
| Browser use is unavailable | A model without tool calling, or a mobile or comparison layout | Switch to a tool-calling model in the regular chat view |
| It describes actions but does not run them | Browser use is off | Ask for one concrete action at a time |
| Part of a page does not work | Embedded content or a captcha | Stop, handle it yourself, then continue |
| A local page will not open | Localhost and file:// are blocked | Move it to a public URL or open it as an attachment |
Where it falls short
Models with shaky tool calling stall halfway, and the more complex the page the more you should avoid asking for several actions at once.
Captchas, embedded widgets and canvas dragging fail easily, so do not use it to automate payments or sign-ins.
It is a feature inside one app, so a UI change changes the steps; running browser automation on a server needs something like Playwright, so this app is not the pick there.
Summarising a page and operating it are different jobs; if you only need a summary, pasting the URL into the chat is faster than handing over the browser.
The app is free but model licences are separate; check the commercial terms on each model card.
Terminology notes
Context window — The token span of input and generated content a model can handle in one request. The supported limit and memory use depend on the model and runtime settings.
Back to the textQuantization — A way to represent model weights or values with fewer bits, reducing storage and memory use. The coarser representation can also affect accuracy.
Back to the textPrompt — The input text sent to a model. For generative models it describes the desired result and strongly shapes the output.
Back to the text
Frequently asked questions
- What should you check first in this article?
- Browser Use hands one browser tab to a local model. The model needs tool calling, and a vision model can also read screenshots. Payments and sign-ins do not proceed without approval, and localhost or file:// addresses do not open.
- How much memory does it need?
- The documentation uses Qwen3.8 27B, but an 8 GB machine of memory should start with a 2B or 4B tool-calling model, and pick a vision model if it has to read images on screen.
- When is it better not to choose this?
- It is a feature inside one app, so a UI change changes the steps; running browser automation on a server needs something like Playwright, so this app is not the pick there.
- Where can you check the conditions and sources?
- The app is free but model licences are separate; check the commercial terms on each model card.