Vercel Agent Browser: A Practical CLI for AI Browser Automation

A coding agent can generate a React page, but it still needs a reliable way to open that page, inspect the controls, fill a form, and confirm that the result works.

That is the niche for Vercel's agent-browser: a command-line browser automation tool designed for AI agents. Instead of requiring an agent to reason over a huge DOM dump or a long list of brittle selectors, it offers a compact workflow built around page snapshots and element references.

A GitHub trend snapshot collected on November 13, 2026 at 06:57 UTC recorded agent-browser among the week's strongest movers, with 4,601 stars gained in the snapshot window. Community interest is not the same as production reliability, though. Here is how the tool works and where I would use it.

The core idea: open, snapshot, interact

The recommended workflow is small:

1. Open a page.
2. Capture an interactive snapshot.
3. Use the returned references to click, fill or read elements.
4. Take a fresh snapshot after the page changes.

Install the CLI globally:

npm install -g agent-browser
agent-browser install

The second command downloads Chrome for Testing the first time. If you only want to try it without a global install:

npx agent-browser install
npx agent-browser open example.com

On Linux, the official installation guide also documents installing browser system dependencies with:

agent-browser install --with-deps

For an AI coding assistant, install the maintained skill instead of copying an old instruction file into your project:

npx skills add vercel-labs/agent-browser

The official docs list integrations for Claude Code, Codex, Cursor, Gemini CLI, GitHub Copilot, Goose, OpenCode and Windsurf.

A first browser workflow

Open a page and inspect its interactive elements:

agent-browser open https://example.com
agent-browser snapshot -i

A snapshot can return references such as:

@e1 [heading] "Example Domain"
@e2 [link] "More information..."

Now interact with the reference:

agent-browser click @e2
agent-browser snapshot -i
agent-browser close

For a form, the same pattern might look like this:

agent-browser open https://example.com/contact
agent-browser snapshot -i
agent-browser fill @e3 "[email protected]"
agent-browser click @e4
agent-browser snapshot -i

The reference IDs are useful because the agent can point at an element it has just observed rather than reconstructing a selector from memory. They are not permanent IDs: after navigation or a meaningful page update, take another snapshot and use the new references.

Why snapshots matter for agents

Traditional browser automation often relies on CSS selectors, XPath, or generated scripts. Those are still supported, but they can be awkward for an LLM to reason about when the page is changing.

Agent-browser also supports semantic locators:

agent-browser find role button click --name "Submit"
agent-browser find label "Email" fill "[email protected]"
agent-browser find placeholder "Search..." fill "browser automation"

The interactive snapshot mode is especially useful when you only need controls that can be clicked or filled. It can reduce the amount of page context the agent has to process.

Other useful capabilities include:

• Headed mode for watching a browser while debugging
• Screenshots for visual verification
• Isolated sessions for separate cookies and browser state
• Chrome DevTools Protocol connections to an existing browser
• Waiting for a specific element, URL or text instead of sleeping for an arbitrary duration
• JSON output for scripts and downstream tools
• Network controls, recording and debugging features

The project uses a fast Rust CLI with a Node.js/Playwright-based browser workflow. The CLI and browser work together; this is not a replacement for the browser engine itself.

Where I would use it

1. Test an AI-generated UI. Ask an agent to build a form, then have it open the local development server, submit invalid values, inspect validation messages and capture a screenshot.

2. Reproduce a user journey. Automate a predictable sequence such as sign-in, navigation and form completion in a test environment.

3. Verify a regression. After an agent changes a page, run the same browser journey and check that the expected controls and content remain present.

4. Give a coding agent a browser tool without building a custom MCP server. A CLI can be easier to inspect and run manually when an agent gets stuck.

For deterministic application tests, Playwright's own test runner is still a strong choice. Agent-browser is most interesting when the operator is an AI agent that needs a small, inspectable command surface.

Trade-offs and reliability notes

There are a few things to plan for:

• Browser installation is a real dependency. The first install downloads Chrome for Testing. Restricted networks, missing Linux libraries or download failures can interrupt setup.
• References can become stale. Re-snapshot after navigation, modal changes, client-side route transitions or major DOM updates.
• Automation is not a guarantee of correctness. A successful click does not prove that a payment, account change or backend operation completed correctly. Verify the resulting state.
• Websites can block automation. Respect terms of service, rate limits, robots policies and account permissions. Do not use browser automation to bypass access controls.
• Choose the right test layer. Use Playwright tests for stable regression suites and agent-browser for agent-driven exploration or task execution.

Community discussions have been mixed in useful ways. Some developers liked the compact snapshot workflow and the ability to guide an agent through a CLI. Others reported browser-download friction and questioned reliability on dynamic applications. Those reports are anecdotal and version-specific, but they are a good reminder to test your own target site before making the tool part of a critical workflow.

Watch a practical walkthrough

This is an independent tutorial, not a Vercel-maintained demo. Use the official quick start for current commands.

Sources and further reading

• Vercel agent-browser repository
• Official installation guide
• Quick start
• Selectors and snapshot references
• Reddit discussion: agent-browser and Claude Code
• Reddit discussion: installation friction
• GitHub weekly trending snapshot for November 13, 2026

What would you automate first with an agent-controlled browser: UI regression checks, form testing, or a repetitive admin workflow?

Written by Sh Raj. Find more engineering notes on GitHub, connect on LinkedIn, and watch practical developer tutorials on StackShade.