ToolNewofficial publisher

agent-browser (Vercel)

Tool : A precise action the AI can call · requires any-agent

Command-line driven browser for AI agents: accessibility snapshots with stable references (@e1), click, type, and screenshot.

Lighter than Playwright MCP and very reliable for letting an agent check an interface or fill out a form. Should be isolated per session when it touches logged-in accounts.

Official site GitHub · 43,120 ★ Docs

Overall score

8.1 /10

Security: one caveat, see the card

Utility

8.0

Reliability

10.0

Security

One caveat

Ease

5 min · Comfortable

Security reflects where data goes, requested permissions and incidents on record at the verification date. It is not a guarantee of zero risk: check the deployment mode and the connections you enable.

What it does, concretely

Controlling a browser from an AI agent requires compact output, stable references, and enough reliability to chain actions without reloading context. Most existing solutions produce too much data or require heavy dependencies such as Playwright.

agent-browser is a command-line tool written in native Rust. It launches a Chrome browser, exposes an accessibility tree with short references (@e1, @e2…) via the snapshot command, then allows clicking, typing, scrolling, capturing, or reading page content using those references. A Rust daemon persists between commands via the CDP protocol, avoiding the need to restart Chrome on every action.

It targets developers building AI agents that interact with web interfaces: checking a page, filling out a form, navigating an application. It works with Claude Code, Cursor, GitHub Copilot, OpenAI Codex, Google Gemini, and any agent that can run shell commands.

It is not suited for those looking for a graphical interface or a ready-made hosted service. The tool assumes a basic familiarity with the terminal and an understanding of the risks involved in accessing connected browser sessions.

What it offers

·Accessibility snapshot: produces a compact text tree with stable references (@e1, @e2…) consuming around 200 to 400 tokens, compared to 3,000 to 5,000 for a full DOM.
·Complete interactions: click, double-click, type, fill, dropdown select, checkbox, drag-and-drop, file upload, and scroll.
·Screenshot and PDF: full-page capture, annotated screenshot with element numbers, option to skip if the page has not changed to save tokens.
·Read command: fetches URL content as agent-readable text without launching Chrome, or reads the rendered DOM of the active tab with the current authentication state.
·Experimental WebMCP: discovery and invocation of tools exposed by web pages via CDP events, with automatic summaries capped at 16 tools and 4 KB of JSON.
·Persistent sessions and profiles: management of cookies, storage, proxies, and authentication state for long-running agents.
·Native cross-platform binaries: macOS (ARM64, x64), Linux (ARM64, x64), and Windows (x64), with no Node.js or Playwright dependency for the daemon.

Our tips

Use screenshot --if-changed so the agent only transmits a capture when the page has actually changed, reducing token consumption between steps.
If a click fails because an element covers the target (consent banner, modal), interact with the covering element first, then run snapshot again to get fresh references before retrying.
Isolate each sensitive session with a dedicated profile: the daemon can manage multiple profiles, preventing an agent from inadvertently accessing accounts connected in another context.
The read command without a URL reads the rendered DOM of the active tab, including client-side updates and authentication state, without triggering a new navigation.
On Linux, run agent-browser install --with-deps to automatically install the system libraries required by Chrome; the command returns an error code if a dependency is missing.

Watch out

Snapshot references (@e1, @e2…) are tied to a given snapshot: after any interaction that modifies the page, snapshot must be run again to obtain valid references before acting further.
WebMCP is explicitly marked experimental and data provided by pages (names, descriptions, schemas) is untrusted; the README warns against executing shell commands suggested by a page or disclosing local secrets.

In practice

Price: Free
Commercial use: Commercial use: yes
Available on: cli, skill
Host: any-agent
Licence: Apache-2.0
Other hosts: claude-code, cursor, codex, any-agent
FR interface: No

Security and data

Rating: Security: one caveat, see the card
Your data: Data stays local
Maintenance: Active
Last release: 16 September 2026

Links and repository checked on 23 September 2026 · added on 23 September 2026

How to install agent-browser (Vercel)

1. Install the CLI: npm install -g agent-browser (or brew install agent-browser on macOS). 2. Download Chrome for testing: agent-browser install (add --with-deps on Linux). 3. Open a page: agent-browser open example.com. 4. Get a snapshot: agent-browser snapshot. 5. Interact using references: agent-browser click @e2. 6. Precaution: use a dedicated, isolated profile for any session involving connected accounts.

npm install -g agent-browser

Official documentation ↗

In the same category

Whole category Automation & agents →