Rank #08 · Daily · Agent-Skill
agent-browser: Vercel Labs' Rust Browser Automation CLI for AI Agents
Browser computer-use for any agent: snapshot returns compact refs like @e1, the agent clicks or fills, then snapshots again — one short shell call per step.
What it does
agent-browser is a native Rust CLI (npm-distributed) that lets an AI agent drive a browser one command at a time. The loop mirrors Vercel Labs' desktop sibling (agent-desktop, #02 on today's list):
- Open a page (
agent-browser open example.com). - Snapshot (
agent-browser snapshot -i) — page as compact accessibility-tree text with a ref per interactive element. - Act on a ref (
click @e1,fill,type, …), then snapshot again to verify.
Design points from the README:
- A background daemon owns the browser over Chrome DevTools Protocol; the browser stays open between commands. Each
--sessiongets its own daemon. - Default browser is Chrome for Testing (
agent-browser install). Also: attach to running Chrome (--cdp/--auto-connect), Lightpanda / Obscura engines, iOS Simulator Safari, and cloud providers (Browserbase, Browserless, Browser Use, Kernel, AgentCore). - Refs are snapshot-scoped; after DOM/navigation changes, take a new snapshot. Clicks fail loudly when the target is hidden, disabled, covered, or off-viewport — dismiss overlays and retry.
- CSS / text / XPath / semantic locators also work (
find role button click --name "Submit"). - Skills ship version-matched with the CLI (
agent-browser skills get core); install discovery via skills.sh or add anAGENTS.mdstub.agent-browser mcpstarts an MCP server over stdio. - Command surface covers navigate/read, interact, wait, capture (screenshot/pdf/record), network/cookies/storage, tabs/frames/dialogs, debugging/profiler, sessions/auth policies, streaming/dashboard, and setup (
install/upgrade/doctor). - Native binaries for macOS / Linux / Windows; docs at agent-browser.dev. License: Apache-2.0.
Who it's for
- Builders of coding agents and MCP tools that need reliable web control without screenshot-coordinate guessing.
- Teams that want a CLI-first, model-agnostic browser tool with allowlists / auth vault / action policies (see Security docs).
- Developers already in the Vercel Labs agent stack who also use agent-desktop for native apps.
Less ideal if you only need a classic Playwright test suite with no agent loop — Playwright remains a fine fit for that.
Stars / language / license
45,893 stars · Rust · Apache-2.0
Data verified: 2026-11-06 09:09 Asia/Shanghai (GitHub API get_repository + README)
Trending note: +150 stars today (Trending selection captured 09:06 Asia/Shanghai)
Category tags
Agent-Skill · browser-automation · computer-use · Rust · CLI · MCP · accessibility · Vercel
How it differs from similar projects
Compared with Playwright / Puppeteer / Selenium. Those are general automation SDKs aimed at tests and scripts. agent-browser is shaped as an agent tool: compact accessibility snapshots, ref handles, shell-sized steps, MCP/skills packaging, and session/security affordances for untrusted agent loops.
Compared with vision computer-use (screenshot → click coordinates). Same trade-off as agent-desktop: structure-first refs vs pixel guessing. Accessibility metadata quality still limits what the agent can see.
Compared with agent-desktop (#02). Same family and loop; agent-browser stops at the browser, agent-desktop drives native OS windows.
Compared with hosted browser-agent products. agent-browser runs locally (or attaches to a cloud browser you choose). You keep the planning model; it only supplies the browser I/O.
TidyHow has not measured click success rates or latency vs Playwright agent wrappers. Treat the differences as architectural.
Minimal loop (from README)
agent-browser open example.com
agent-browser snapshot -i
agent-browser click @e1
agent-browser get url
agent-browser close