Vercel Labs has open-sourced Agent Browser, a command-line tool. Its goal is straightforward: to let AI not only understand webpages but also interact with them.


This isn’t a browser for humans. It’s a low-level interface specifically designed for AI agents. You can think of it as equipping ChatGPT or Claude with hands that can click, type, and scroll through webpages.

Traditional tools like Playwright or Selenium are designed for human developers, requiring scripting and handling complex selectors. Agent Browser takes a different approach, with its core being the **Snapshot + Ref mechanism**.

The AI first executes the `agent-browser snapshot` command to obtain a list of interactive elements on the webpage, each with a unique reference, such as `@e1` (login button), `@e2` (email input box). Then, the AI directly operates using these references: `agent-browser click @e1`. This method is more friendly to AI, avoiding the instability of traditional CSS selectors and making execution more certain and faster.

The tool has rewritten the CLI layer in Rust, making it faster than Node.js/Python solutions. Installation is simple: `npm install -g agent-browser`. It supports two modes: UI-based debugging and silent execution without a UI.


Vercel also provides Claude Code skill templates. Developers place the templates in the `.claude/skills/` directory, and Claude can directly invoke Agent Browser to execute webpage tasks without additional adaptation.

Some users mentioned that combining it with large models like deepseek 671B Code should be very smooth. AI operations on webpages eliminate intermediate steps, allowing tasks to be completed directly.

The emergence of Agent Browser marks the shift of AI from ‘dialogue-based intelligence’ to ‘action-based intelligence’. It may become the infrastructure for future AI workflows, bringing automation to fruition.

Project link: [https://github.com/vercel-labs/agent-browser](https://github.com/vercel-labs/agent-browser)