Skip to content
ReviewRamble

The weekly brief

AI, reviewed, in your inbox.

  • New reviews of tools, models, MCP servers and repos, with the verdict up front
  • Re-checked prices and ratings when something changes
  • New analysis pieces as they are published

Models, agents, MCP servers and the tools actually worth your time.

One email a week. Unsubscribe with the link in any brief. How we use your email

agent-browser

Vercel Labs' Rust CLI that lets coding agents drive Chrome one shell command at a time, reading pages as text with element refs.

Reviewed by ReviewRamble editors · Updated

View on GitHub Open source

Share this page:XLinkedInFacebookRedditWhatsAppEmail

Our rating

4.4/5

The browser CLI most coding agents should reach for first, provided you switch on the guard rails it leaves off by default.

What is agent-browser?

agent-browser is a native Rust command-line tool from Vercel Labs that gives an AI agent a browser to drive. agent-browser snapshot returns the page as compact accessibility text with a ref such as @e1 for each element; the agent clicks or fills by ref, reads the page again and repeats. A background daemon starts on the first command, controls the browser over the Chrome DevTools Protocol and stays open between calls, so each step is one short shell command.

Beyond navigation and input it covers screenshots, PDFs, recordings, visual diffs, cookies and storage, network inspection, console errors, traces, Core Web Vitals and a React inspector. Chrome for Testing is the default; it can also attach to your running Chrome, switch to the Lightpanda or Obscura engines, drive Mobile Safari in the iOS Simulator, or use cloud browsers from Browserbase, Browserless, Browser Use, Kernel and AgentCore. Skills ship with the binary, and agent-browser mcp starts a stdio MCP server.

The repo is Apache-2.0 with about 53,000 GitHub stars; v0.43.0 shipped on 12 November 2026, and the npm package was downloaded about 2.6 million times in the week to 19 November.

Is agent-browser worth it?

agent-browser bets that a coding agent already knows how to use a terminal, so the browser should be a terminal command too. Each step is open, snapshot, click @e3, snapshot again, and the output is a few lines of text rather than a screenshot or a page of HTML. A click that lands on a hidden, disabled or covered element fails instead of reporting success, which saves the agent from confidently clicking a cookie banner for ten turns.

That is the main difference from Playwright MCP, which presents the browser as MCP tools whose definitions the client loads up front; here MCP is an extra, and any agent that can run a shell can use it with no client support at all. It differs from browser-use in a different way: browser-use is a Python library with its own model loop, while agent-browser has no model inside and leaves the thinking to Claude Code, Codex or whichever agent calls it. Its install numbers, about 2.6 million npm downloads in a week, say the approach has caught on.

The catch: out of the box it puts no limits on where the agent goes, what it runs or what it prints. The documented controls are good but opt-in: a domain allowlist, an action policy over 14 command categories, human confirmation that auto-denies after 60 seconds, and nonce-tagged content boundaries to mark page text as untrusted. The auth vault stores logins encrypted with AES-256-GCM so the model never sees the password, but saved session files in ~/.agent-browser/sessions/ hold tokens in plaintext unless you set an encryption key. --profile copies your real Chrome cookies into the session, and --auto-connect needs Chrome's debugging port open, which the docs warn gives any local process full control. The allowlist cannot be combined with either mode.

Our take: for an agent that writes and checks web code, this is the default browser tool now, and the first job after installing it is a launch line with --allowed-domains and an --action-policy file that denies eval. Keep it away from your everyday Chrome profile unless you would hand the agent your logged-in sessions anyway.

Where it earns its place

  • A front-end developer working with Claude Code on a Next.js app

    Have the agent check its own UI change in a browser before saying it is done

    Install with npm install -g agent-browser and agent-browser install, add two lines to AGENTS.md pointing at agent-browser skills get core, then let the agent open localhost, snapshot, click through the flow and read console errors. Free; runs on the laptop's Chrome for Testing.

  • A QA engineer at a 20-person SaaS company

    Catch layout regressions and slow pages on a staging site

    Script agent-browser diff and vitals against staging in CI, with --allowed-domains set to the staging host and an action policy that denies eval and downloads. No licence cost; cloud browsers from Browserbase or Kernel are optional and billed by them.

  • An operations analyst

    Let an agent pull a monthly export from a portal that needs a login

    Save the login with agent-browser auth and --password-stdin so the model only sees the profile name, set AGENT_BROWSER_ENCRYPTION_KEY so session files are encrypted, and run with --confirm-actions so downloads need a yes.

Before you commit

Setup
npm install -g agent-browser (or brew install agent-browser), then agent-browser install to download Chrome for Testing; on Linux add --with-deps. npx agent-browser open example.com works without installing. Native binaries for macOS, Linux and Windows.
Integrations
Shell CLI, bundled skills (agent-browser skills get core), skills.sh install, stdio MCP server, plugins; engines Chrome for Testing, Lightpanda, Obscura, iOS Simulator; cloud providers Browserbase, Browserless, Browser Use, Kernel and AgentCore; guides for Vercel Sandbox.
Data and privacy
Runs locally. Auth vault in ~/.agent-browser/auth/ is always AES-256-GCM encrypted with an auto-generated key file; session state in ~/.agent-browser/sessions/ is plaintext unless AGENT_BROWSER_ENCRYPTION_KEY is set, and expires after 30 days by default. No published security advisories.
Cost at scale
No paid edition. Cost is the machine running Chrome plus model tokens, and per-minute fees if you choose a cloud browser provider.
Limits
Refs go stale after navigation or DOM changes and need a fresh snapshot. The allowlist is best-effort, not a firewall. The daemon closes a headless browser after one hour idle by default. Version 0.43.0 is still pre-1.0.

What’s good

  • Compact text snapshots with refs keep each step to a few lines
  • Plain shell commands work with any agent; MCP optional
  • Wide coverage: network, traces, vitals, diffs, recording, iOS Simulator
  • Encrypted auth vault keeps passwords out of the model's context

What’s not

  • No navigation, action or output limits unless you configure them
  • Saved session state is plaintext without an encryption key
  • Domain allowlist cannot be used with --profile, --cdp or --auto-connect
  • Pre-1.0 and fast-moving; command surface changes between releases

Best for

Developers who want Claude Code, Codex or another terminal agent to test and debug web apps in a real browser with little context overhead.

agent-browser alternatives

agent-browser: common questions

Is agent-browser worth it?
The browser CLI most coding agents should reach for first, provided you switch on the guard rails it leaves off by default. We rate it 4.4 out of 5.
How much does agent-browser cost?
Free and open source (Apache-2.0). Optional cloud browser providers bill separately.
What are the downsides of agent-browser?
No navigation, action or output limits unless you configure them; saved session state is plaintext without an encryption key; domain allowlist cannot be used with --profile, --cdp or --auto-connect; pre-1.0 and fast-moving; command surface changes between releases.
Who is agent-browser best for?
Developers who want Claude Code, Codex or another terminal agent to test and debug web apps in a real browser with little context overhead.
How much does agent-browser cost for a team or in production?
No paid edition. Cost is the machine running Chrome plus model tokens, and per-minute fees if you choose a cloud browser provider.
Is agent-browser safe for company data?
Runs locally. Auth vault in ~/.agent-browser/auth/ is always AES-256-GCM encrypted with an auto-generated key file; session state in ~/.agent-browser/sessions/ is plaintext unless AGENT_BROWSER_ENCRYPTION_KEY is set, and expires after 30 days by default. No published security advisories.
What are the best agent-browser alternatives?
The closest alternatives we have reviewed are Browser Use, Stagehand, PinchTab and Lightpanda.

Answers checked Nov 20, 2026.

Make agent-browser? Show this rating on your siteagent-browser rated 4.4 out of 5 on ReviewRamble