AI GitHub
Browser Use for model-driven web tasks
Let a model navigate a browser and act on page elements.
Key Takeaways
- The agent receives page observations and chooses browser actions within a task loop.
- Choose a browser session and a model with the necessary input support.
- Test a navigation task on a disposable site with two similar buttons.
- Webpage text can contain instructions aimed at the agent.
How it works
The agent receives page observations and chooses browser actions within a task loop. Browser and model configuration determine what it can see and which sessions it can use.
What you need
Choose a browser session and a model with the necessary input support. Use a test account whenever a workflow needs to see an authenticated page.
What to try first
Test a navigation task on a disposable site with two similar buttons. Save the action history and check the actual page state at completion.
Record the URL and selected element after each action. The final answer alone may claim success even when a modal or validation screen interrupted the task.
Where to look for failures
Webpage text can contain instructions aimed at the agent. Keep valuable sessions and irreversible actions outside the initial test.
Related Franklin pages
Hold the browser task constant and compare the observation and action layers. An explicit selector, a structured browser snapshot and a model-chosen action fail in different ways.
- Combine explicit browser steps with natural-language actions, extraction and observation. Stagehand for code and model-assisted browser automation
- Expose browser navigation and interaction through a Model Context Protocol server. Playwright MCP for agent browser tools
For an agent interface that reaches mobile apps, see Franklin’s guide to Phone-Use. Phone-Use
Project documentation
Try it for your workflow
Explore the project’s documentation for setup and usage, or save this page to a collection for later.
Read the project documentation ↗