Browser

Drives a headless browser from the command line through agent-browser, a Rust CLI daemon. Opens pages, clicks, fills forms, takes screenshots, runs JS, intercepts network, and emulates devices. Holds per-site auth profiles so you log in once and stay logged in for headless runs after that. Batches commands, runs sessions in parallel, and hands browser work to background agents.

Most browser automation is slow to script, leaks state between runs, and forces a fresh login every time. You either babysit a headed browser or fight a framework that re-authenticates on every call. When you need to screenshot ten pages, scrape a logged-in site, or run several scrapes at once, the per-run startup and auth cost dominates. This skill keeps a daemon warm, persists auth per site, and lets you fan work out to parallel sessions and agents.

Tool: agent-browser — headless Rust CLI daemon with persistent auth profiles.

If agent-browser isn't working or a site has bot detection, use the Interceptor skill instead. Interceptor is a Chrome extension with zero CDP fingerprint — passes all major bot detection checks.

Does the site need auth?

Use --profile ~/.agent-browser/profiles/<site>. If profile exists, auth is automatic. If not, run --headed once for login, then headless forever.

Native Rust daemon. Persistent profiles for auth. Headless by default.

Quick One-Shot Commands

agent-browser open https://example.com && agent-browser screenshot /tmp/shot.png
agent-browser open https://example.com && agent-browser screenshot --full /tmp/full.png
agent-browser open https://example.com && agent-browser pdf /tmp/page.pdf

Session-Based Interaction

# 1. OPEN
agent-browser open https://example.com

# 2. WORK
agent-browser snapshot                    # a11y tree with @eN refs (for AI)
agent-browser click @e12                  # click by ref
agent-browser fill @e15 "hello"           # fill input by ref
agent-browser screenshot /tmp/shot.png    # screenshot
agent-browser eval "document.title"       # run JS

# 3. CLOSE — when done
agent-browser close

Authenticated Browsing (Per-Site Profiles)

First-time setup (headed, one-time):

# Close any running daemon first
agent-browser close --all

# Launch headed with persistent profile — log in manually
agent-browser --headed --profile ~/.agent-browser/profiles/<site> open https://example.com

# After login completes, all future runs reuse the profile headlessly

Subsequent runs (headless, automatic):

agent-browser --profile ~/.agent-browser/profiles/<site> open https://example.com
# Auth is automatic — cookies, IndexedDB, cache all persist

To add a new site: Close daemon, run --headed --profile ~/.agent-browser/profiles/<name> once, log in, done.

Auth Vault (Alternative)

agent-browser auth save mysite --url https://example.com --username user --password-stdin
agent-browser auth login mysite    # auto-fills login form
agent-browser auth list            # show saved profiles

Batch Execution

# Send multiple commands in one shot (fewer tool calls = fewer tokens)
echo '[["open","https://example.com"],["snapshot"],["click","@e12"]]' | agent-browser batch

Advanced Features

# Connect to already-running Chrome
agent-browser --auto-connect snapshot

# Network interception
agent-browser route "**/*.{png,jpg}" abort     # block images
agent-browser route "https://api.com/*" mock '{"data":"test"}'

# Device emulation
agent-browser --device "iPhone 15" open https://example.com

# Session persistence (cookies + localStorage by name)
agent-browser --session-name myapp open https://example.com

agent-browser Rules

• Daemon model — first command starts daemon, subsequent commands connect instantly.
• Refs use @eN syntax — @e12 not e12.
• Profiles persist everything — cookies, IndexedDB, cache, localStorage.
• Close with agent-browser close or close --all to kill daemon.

Delegating Browser Work to Agents

When you need parallel or background browser work (scraping multiple pages, monitoring), spawn general-purpose agents with browser instructions. No dedicated browser agent type needed — this skill IS the expertise.

Agent(subagent_type="general-purpose", prompt="
  Use agent-browser CLI for all browser work.
  Commands: open <url>, snapshot, click @eN, fill @eN 'text', screenshot /path.
  For authenticated sites: --profile ~/.agent-browser/profiles/<site>
  Refs use @eN syntax from snapshots.
  [your specific task instructions here]
")

For parallel isolation, each agent uses --session <name>:

Agent 1: agent-browser --session scrape1 open https://site-a.com
Agent 2: agent-browser --session scrape2 open https://site-b.com

Fallback: If agent-browser fails or the site has bot detection, use the Interceptor skill instead.

Legacy built-in agents — REMOVED 2026-06-10. BrowserAgent and UIReviewer were Claude Code built-ins whose internals cannot be modified; they run browser automation that PAI no longer uses. Route all browser work through the Interceptor skill (verification, authenticated flows) or agent-browser (headless scraping).

• SKIP-gate (check before anything else): task is deploy-verification or UI confirmation → STOP, route to Interceptor. Machine-checkable precheck: scan the request for deploy|verify|verification|confirm.*UI|UI.*confirm — if the task is verifying a deploy or confirming a UI change, the PAI_SYSTEM_PROMPT mandates Interceptor (real Chrome), not headless Browser. When this gate fires, announce the skip and why in the response ("Skipping Browser — deploy/UI verification routes to Interceptor per system-prompt mandate") — a silent skip is a failure.

Define user stories in YAML and validate them in parallel with general-purpose reviewer agents.

Directory: skills/Browser/Stories/

name: App Name
url: https://example.com
stories:
  - name: Story name
    steps:
      - action: click
        target: "LLM-readable description"
    assertions:
      - type: snapshot_contains
        text: "expected text"

Run with: "review stories" or "run stories in HackerNews.yaml"

Reusable Markdown templates with {PROMPT} injection.

Directory: skills/Browser/Recipes/

Run with: "automate SummarizePage for https://example.com"

Example 1: Batch screenshots of a dev server

User: "screenshot the five main pages of the dev server"
→ agent-browser open + screenshot per page (batch mode, one daemon)
→ Saves PNGs, returns paths

Example 2: Authenticated scrape

User: "extract the data table from my dashboard"
→ Opens with --profile ~/.agent-browser/profiles/<site> (auth persists from one-time headed login)
→ snapshot + eval to pull the table, returns structured text

Example 3: Story validation

User: "review stories in HackerNews.yaml"
→ ReviewStories workflow: fans YAML stories out to parallel reviewer agents
→ Each agent runs steps + assertions, results aggregated