What problem it solves

AI agents need a way to open web pages, understand their contents, and interact with page elements through a simple interface. Repeated browser actions also require stable element references and a browser session that remains available across commands.

Key features

• Reference-based browser interaction — Snapshots represent page elements as compact text with references that agents can use for subsequent actions. References are refreshed after navigation or DOM changes.

  Source: README

• Persistent browser sessions — A background daemon controls the browser over the Chrome DevTools Protocol and keeps the browser open between commands. Each session can use its own daemon and browser.

  Source: README

• Multiple interaction and inspection commands — The command set covers navigation, reading, interaction, waiting, capture, network and page state, sessions, debugging, and integrations. It also supports CSS selectors, text, XPath, and semantic locators.

  Source: README, cli

• Agent skills and MCP support — Version-matched skills provide the core interaction loop, command references, templates, and specialized workflows. An MCP server can also expose the CLI to MCP clients over stdio.

  Source: README

• Flexible browser connections — Chrome for Testing is the default, while the CLI can attach to an existing Chrome, select another supported engine, or use listed cloud browser providers.

  Source: README

Quick start

Copied verbatim from the project's README.

npm install -g agent-browser
agent-browser install           # download Chrome for Testing (first time only)

agent-browser open example.com
agent-browser snapshot -i       # interactive elements, each with a ref
agent-browser click @e1         # act on a ref from that snapshot
agent-browser get url           # check the result
agent-browser close

Who it may suit

Inference: This may suit developers building AI-agent workflows that need browser interaction from the shell or an MCP client.

Topics to discuss

• Which browser-engine and provider combinations should be tested most extensively?

• How can agents handle stale references and changed page structures more reliably?

• Which skill or MCP workflows would be most useful in real agent harnesses?

Written by ReadmeX AI from the project's README, source and public GitHub data; it may contain mistakes.

Read the original README https://github.com/vercel-labs/agent-browser#readme

ReadmeX opened this community for a trending project; it is not run by the project. Maintainers can claim it from the community page.