S3YED agent-browser — fleet default for browser automation
A native-Rust browser-automation CLI designed for AI agents. Compact text output minimises context usage. Use it instead of heavier browser stacks for shell-driven web tasks.
Install
npm install -g agent-browser # all platforms
# or: brew install agent-browser # macOS
agent-browser install # download Chrome (first run only)
# try without installing:
npx agent-browser open example.com
Core loop (ref-based)
agent-browser open example.com
agent-browser snapshot -i # accessibility tree with refs: heading "..." [ref=e1], link "..." [ref=e2]
agent-browser click @e2 # act on a ref from the snapshot
agent-browser screenshot page.png
agent-browser close
Why refs (the context win)
snapshot returns a compact accessibility tree where each element has a ref (@e1, @e2):
- Context-efficient: ~200-400 tokens vs ~3000-5000 for full DOM.
- Deterministic: a ref points to the exact element from the snapshot (no fragile selectors).
- Fast: no DOM re-query. AI-friendly: text output parses naturally.
What it covers (50+ commands)
Navigation, forms, screenshots, network control, storage/cookies/auth-state, file upload/download, tabs, frames, debugging. Built-in: video recording, streaming, profiler, diffing. First-class docs for React/Web Vitals, init scripts, and Next.js + Vercel. Stateful: sessions, profiles, proxy, security controls — good for long-running agents.
React DevTools introspection (v0.27+)
Launch a browser with React DevTools pre-loaded:
agent-browser open --enable react-devtools http://localhost:3000
React DevTools CLI (v0.27+)
Launch with --enable react-devtools for React component introspection:
| Command | Description |
|---|---|
react tree |
Print React component tree with fiber IDs |
react inspect <id> |
Inspect props, hooks, state, source for one component |
react renders start/stop |
Profile re-renders |
react suspense --only-dynamic --json |
Walk Suspense boundaries, classify static vs dynamic |
Core Commands
| Command | Description |
|---|---|
react tree |
Print the React component tree with fiber IDs |
react inspect <id> |
Inspect props, hooks, state, and source for one component |
react renders start |
Begin recording component render commits |
react renders stop [--json] |
Stop recording and print a render profile |
react suspense [--only-dynamic] [--json] |
Walk Suspense boundaries, classify static vs dynamic |
Typical workflow:
# Open with React DevTools
agent-browser open --enable react-devtools http://localhost:3000
# Inspect component tree
agent-browser react tree
# Inspect a specific component
agent-browser react inspect 42
# Profile re-renders after an interaction
agent-browser click @e3
agent-browser react renders stop
# Check Suspense boundaries
agent-browser react suspense --only-dynamic
If the session already has a browser without the hook, pass --enable react-devtools to relaunch with it installed.
Web Vitals
Measure Core Web Vitals on any site:
agent-browser vitals
agent-browser vitals https://example.com --json
Reports LCP, CLS, TTFB, FCP, INP, and hydration timing when available (React profiling build).
SPA navigation
Client-side navigation without full reload:
agent-browser pushstate /dashboard
agent-browser wait --load networkidle
agent-browser snapshot -i
On Next.js apps, uses window.next.router.push so RSC fetches still run. Other frameworks fall back to history.pushState.
Integration with Next.js 16.3
Combined with the next-dev-loop skill (Next.js 16.3+), agent-browser gives agents the full dev feedback loop:
# Install the skill
npx skills add vercel/next.js --skill next-dev-loop
# In your agent prompt: "After every edit, verify the page still works at runtime using the next-dev-loop skill."
Architecture
Client-daemon: a Rust CLI talks to a native Rust daemon that drives Chrome over CDP. The daemon starts automatically and persists between commands.
When to use vs other tools
- Default for shell-driven browser work (cheapest context, deterministic).
- Use Playwright/Stagehand only when you need their specific ecosystem features.
- Cross-platform: macOS (arm64/x64), Linux (arm64/x64), Windows (x64).
Works with: Claude Code, Cursor, Copilot, Codex, Gemini, opencode, and any agent that runs shell commands.