CLI Overview

CLI-first browser automation for AI coding agents. Like playwright-cli but with composite actions, semantic selectors, and structured extraction.

BAP CLI

CLI-first browser automation from the command line. BAP defaults to installed Chrome, an auto-detected profile when available, and a persistent daemon so agents can work against a real browser instead of starting fresh every time.

Install

npm

npm i -g @browseragentprotocol/cli

Requires Node.js 20+ and Playwright browsers. Install browsers with npx playwright install chromium.

30-Second Demo

1. Navigate to a page

   bap goto https://example.com --observe
   

   The --observe flag fuses navigation and observation into a single server call, returning interactive elements with stable refs.

2. Click an element

   bash bap click role:button:"Get Started" Semantic selectors target elements by purpose, not DOM position. They survive redesigns.

3. Fill a form with composite actions

   bap act fill:label:"Email"="[e-mail]" \ fill:label:"Password"="secret" \

   click:role:button:"Sign in" ``` Three steps, one command, one server call. Fewer tokens, fewer
   roundtrips.
   </Step>
   
   <Step title="Extract structured data">
     ```bash
     bap extract --fields="title,price,rating"
   

   Returns validated JSON output.

Architecture

bap <command>
    | WebSocket (JSON-RPC 2.0)
BAP Playwright Server (auto-started as background daemon)
    | Playwright
Browser (Chromium / Firefox / WebKit)

The server starts automatically on first use and persists across commands. No manual setup needed.

Use bap close-all to stop the daemon and shut down all browser sessions.

Why BAP CLI?

Composite Actions

Execute multi-step flows in one command instead of one-at-a-time. A login flow is 1 bap act, not 3 separate fill/click calls.

Semantic Selectors

Target elements by ARIA role, visible text, label, or placeholder. Selectors survive DOM changes and redesigns.

Structured Extraction

Extract validated JSON from pages using field names, JSON Schema, or list mode. No HTML parsing needed.

Composite Actions vs playwright-cli

# playwright-cli: 3 commands, 3 snapshots, 3 LLM reasoning cycles
playwright-cli fill e5 "[e-mail]"
playwright-cli fill e8 "password123"
playwright-cli click e12

# bap: 1 command, 1 snapshot, 1 LLM reasoning cycle
bap act fill:e5="[e-mail]" fill:e8="password123" click:e12

Migrating from playwright-cli

BAP is a drop-in replacement. All e<N> refs from snapshots work identically:

playwright-cli	bap
playwright-cli open [url]	bap open [url]
playwright-cli click e15	bap click e15
playwright-cli fill e5 "text"	bap fill e5 "text"
playwright-cli snapshot	bap snapshot
playwright-cli screenshot	bap screenshot

BAP adds composite actions, semantic selectors, smart observation, and structured extraction on top.

Output

Commands produce concise, AI-agent-friendly output. Files are saved to .bap/ in the current directory:

• Snapshots: .bap/snapshot-<timestamp>.yml
• Screenshots: .bap/screenshot-<timestamp>.png
• Extractions: .bap/extraction-<timestamp>.json

Command Overview

Category	Commands
Essentials	goto, observe, act, extract, screenshot
Actions	click, fill, type, press, select, check/uncheck, hover, scroll
Navigation	open, back, forward, reload, close, close-all
Observation	observe, snapshot
Sessions & Tabs	sessions, tabs, tab-new, tab-select, frames, frame-switch
Workflows	workflow record/stop/run/list, run
Debugging	watch, trace, demo, plan
Configuration	config, install-skill, skill