Agent Browser Loop

Let your coding agent verify its own work.

AI coding agents can write code, run type checks, even execute unit tests - but they can't click through the app like a user would. They get stuck waiting for humans to manually verify "does the button work? does the form submit? does the error message appear?"

Agent Browser Loop gives agents a browser they can drive. Write code, navigate to the page, fill the form, click submit, check the browser logs, take a screenshot, see the error, fix it, retry - all without human intervention.

This doesn't eliminate human review - but it lets agents verify and unblock themselves instead of stopping at every turn. Engineers can give agents a longer leash, and agents can build features end-to-end while proving they actually work.

---

Install

Requires Bun.

bun install -g agent-browser-loop
agent-browser setup

This installs the CLI globally, downloads Playwright Chromium, and copies skill files to .claude/skills/ so Claude, OpenCode, and other AI agents know how to use the browser.

Quick Start

agent-browser open http://localhost:3000 --headed
agent-browser act click:button_0 type:input_0:"hello"
agent-browser wait --text "Success"
agent-browser state
agent-browser close

The --headed flag shows the browser. Omit for headless mode.

How Agents Use It

The agent works in a loop: open -> act -> wait/verify -> repeat

# 1. Open the app
agent-browser open http://localhost:3000/login

# 2. Fill form and submit
agent-browser act type:input_0:user@example.com type:input_1:password123 click:button_0

# 3. Wait for navigation
agent-browser wait --text "Welcome back"

# 4. Verify state
agent-browser state

Every command returns the current page state - interactive elements, form values, scroll position, console errors, network failures. The agent sees exactly what it needs to verify the code works or debug why it doesn't.

Profiles (Session Storage)

Save and reuse login sessions across runs:

# Save current session to a profile
agent-browser profile save admin

# Use profile (auto-saves updated tokens on close)
agent-browser open http://localhost:3000 --profile admin

# Use --no-save for read-only access
agent-browser open http://localhost:3000 --profile admin --no-save

# Other commands
agent-browser profile list
agent-browser profile capture admin --url http://localhost:3000/login
agent-browser profile delete admin

Profiles store cookies and localStorage. Use --global for user-level profiles, --private for gitignored project profiles.

Multi-Session

Run multiple browser sessions in parallel:

# Create sessions with auto-generated IDs
agent-browser open --new http://localhost:3000     # Output: Session: swift-fox
agent-browser open --new http://localhost:3000     # Output: Session: calm-river

# Target specific sessions
agent-browser act -s swift-fox click:button_0
agent-browser state -s calm-river

# List all sessions
agent-browser sessions

CLI Reference

Command	Description
open <url>	Open URL (starts daemon if needed)
act <actions...>	Execute actions
wait	Wait for condition
state	Get current page state
screenshot	Capture screenshot
sessions	List all active sessions
close	Close session or daemon
setup	Install browser + skill files

Actions

agent-browser act click:button_0           # Click element
agent-browser act type:input_0:hello       # Type text
agent-browser act press:Enter              # Press key
agent-browser act scroll:down:500          # Scroll
agent-browser act navigate:http://...      # Navigate

Multiple actions: agent-browser act click:input_0 type:input_0:hello press:Enter

Wait Conditions

agent-browser wait --text "Welcome"        # Text appears
agent-browser wait --selector "#success"   # Element exists
agent-browser wait --url "/dashboard"      # URL matches
agent-browser wait --not-text "Loading"    # Text disappears
agent-browser wait --timeout 60000         # Custom timeout

Options

--headed              # Show browser window
--new                 # Create new session with auto-generated ID
--session <id>        # Target specific session (from --new)
--json                # JSON output
--no-state            # Skip state in response

State Output

URL: http://localhost:3000/login
Title: Login
Scroll: 0px above, 500px below

Interactive Elements:
  [0] ref=input_0 textbox "Email" (placeholder="Enter email")
  [1] ref=input_1 textbox "Password" (type="password")
  [2] ref=button_0 button "Sign In"

Errors:
Console: [error] Failed to load resource: 404
Network: 404 GET /api/user

Use ref values in actions: click:button_0, type:input_0:hello

Screenshots

agent-browser screenshot -o screenshot.png       # Save to file
agent-browser screenshot --full-page -o full.png # Full scrollable page
agent-browser screenshot                         # Output base64

Useful for visual debugging when text state isn't enough to diagnose issues.

HTTP Server Mode

For multi-session scenarios or HTTP integrations:

agent-browser server --headed
# Runs at http://localhost:3790
# API spec at GET /openapi.json

Configuration

CLI flags or config file (agent.browser.config.ts):

import { defineBrowserConfig } from "agent-browser-loop";

export default defineBrowserConfig({
  headless: false,
  viewportWidth: 1440,
  viewportHeight: 900,
});

On macOS, headless system Chrome can crash during AppKit startup. By default, the CLI falls back to bundled Playwright Chromium when headless: true. If you explicitly want system Chrome in headless mode, set allowSystemChromeHeadless: true.

What This Is NOT For

This tool is for agents to test their own code. It is not for:

• Web scraping
• Automating third-party sites
• Bypassing authentication

Use it on your localhost and staging environments.

License

MIT