A browser automation CLI published by Vercel (vercel-labs). It lets an AI agent open a web page, click, fill in forms, and take screenshots ─ all through single-line commands. Its defining trait is a design tuned for output that AI agents read: instead of verbose JSON it returns compact text, and elements are addressed with short identifiers (refs) such as @e1. It pairs with Claude Code, Cursor, GitHub Copilot, OpenAI Codex, Gemini, and any other agent that can run shell commands. Since its release in January 2026, it has shipped close to 50 releases through the end of August.
Key Features
- Ref-based element selection: Running the
snapshotcommand returns the page structure with a short ID assigned to each element, such as@e1or@e2. From then on you simply writeclick @e1, so there is no need to have the AI produce long CSS selectors or XPath expressions. Targets break less often, and the token volume of each exchange stays low - More than 50 commands: Navigation, form input, screenshots, PDF export, JavaScript evaluation, network interception, cookie and storage handling, tab and frame control, and accessibility audits (axe-core, with WCAG tag filtering) are all covered
- A Rust CLI with a daemon: Rather than relaunching the browser on every command, a resident daemon keeps the browser alive over the Chrome DevTools Protocol. The daemon stops automatically after one hour of idle time and saves its state
- Persistent sessions: Cookies, login state, profiles, and proxy settings can be retained, so long-running tasks against sites that require a login do not have to re-authenticate every time
- Runs as an MCP server too:
agent-browser mcpstarts a Model Context Protocol stdio server. You can pick a tool profile ─ core, network, state, debug, tabs, react, mobile, or all ─ to limit how many tools are exposed to the agent - Cloud browsers and iOS Simulator support: The
-poption switches execution to remote browsers such as Browserless, Browserbase, Browser Use, Kernel, or AgentCore. Browsers on the iOS Simulator can also be driven via Appium
Pricing
| Plan | Price | What it covers |
|---|---|---|
| Open source (Apache-2.0) | $0 | All functionality, including the CLI, the daemon, and the MCP server. Runs locally |
| Cloud browser integrations | Each provider’s own pricing | Using Browserbase, Browserless, and similar services incurs separate charges from those vendors |
Pricing reflects information as of August 2026. Check the official site for the latest details.
Pros & Cons
✅ Pros
- Free and open source (Apache-2.0); you can try it immediately without signing up (
npx agent-browser open example.com) - Output is compact text, so it puts less pressure on the agent’s context window
- Ref-based targeting avoids the classic failure mode where AI-generated selectors break
- Usable as both a CLI and an MCP server, so it works with any agent that can reach a shell
- Native binaries are distributed for macOS, Linux, and Windows, and it installs via npm, Homebrew, or Cargo
⚠️ Cons
- It is a command-line tool rather than a GUI application, so getting started is harder without terminal experience
- Development moves very fast (roughly 50 releases between January and August 2026), so behavior differences between versions need attention
- Chrome is required in the execution environment (
agent-browser installis provided to set up Chrome for Testing) - When using remote browsers or cloud providers, their pricing and terms must be tracked separately
- As an open source project, it offers no commercial support desk
Comparison with Similar Services
| Criteria | agent-browser | Playwright MCP | browser-use | Browserbase |
|---|---|---|---|---|
| Provider | Vercel (vercel-labs) | Microsoft | browser-use | Browserbase |
| Form factor | CLI + MCP server | MCP server | Python library | Cloud browser platform |
| Typical usage | Driven directly by shell commands | Attached to an agent as tools | Controlled from Python code | Used as a remote execution environment |
| Where it runs | Local (switchable to remote) | Local | Local or cloud | Cloud |
| Price | Free (open source) | Free (open source) | Free (open source) | Paid |
Who Is It For
- Developers who want Claude Code or Cursor to check and operate web screens as well
- People who want to delegate E2E testing or form-filling automation to an agent instead of writing code
- Anyone who wants to connect browser control tools to an agent through an MCP server
- Teams whose existing automation scripts keep breaking when selectors change
- Front-end engineers who want to automate display checks on mobile environments, including the iOS Simulator
Summary
agent-browser is a tool designed around browser automation that AI reads, rather than browser automation that people read. Ref-based element selection and compact text output let an agent operate a screen reliably with fewer exchanges. It is Apache-2.0 open source and costs nothing, so start by running npx agent-browser open example.com to see how it behaves; if you plan to use it regularly, install it with npm or Homebrew and connect it to your agent as an MCP server.