brow - agentic browsing

Standalone Playwright CLI for agent browser automation. Launches a real Chromium instance with an agent-friendly API - structured commands for common actions, eval escape hatch for full power.

Documentation →

Install

Homebrew:

brew tap detrin/tap
brew install brow

pip:

pip install brow-cli

Then install Chromium once (either method above):

brow setup            # ~150MB, one-time

Update patchright and its matching Chromium build later with:

brow setup --upgrade

Agent skill:

npx -y skills add detrin/brow

Example: Find Bars Near Times Square with Google Maps

A real use case: use your Google account to search Maps in a city you've never visited, and extract structured results.

Step 1: Log into Google (once)

Open a headed browser with a persistent profile and sign in manually:

brow session new --profile personal --headed
brow navigate -s 1 "https://accounts.google.com"
# Sign in manually in the browser window...
brow session delete 1

Your login is saved in ~/.brow/profiles/personal/ -you won't need to sign in again.

Step 2: Ask Claude Code to search

Paste this into Claude Code:

Open a brow session with my personal profile, go to Google Maps, and search for bars near Times Square in New York. Return the names, Google Maps URLs, ratings, and number of reviews in a markdown table.

Claude Code runs:

brow session new --profile personal --headed    # → 1 (already logged in)
brow navigate -s 1 "https://www.google.com/maps/search/bars+near+Times+Square+New+York"
brow screenshot -s 1
brow eval -s 1 "
results = await page.evaluate('''() => {
    const items = document.querySelectorAll('div.Nv2PK');
    return Array.from(items).slice(0, 8).map(el => {
        const name = el.querySelector('.fontHeadlineSmall, .qBF1Pd');
        const rating = el.querySelector('.MW4etd');
        const reviews = el.querySelector('.UY7F9');
        const link = el.querySelector('a[href*=\"/maps/place\"]');
        return {
            name: name?.innerText || '',
            rating: rating?.innerText || '',
            reviews: reviews?.innerText.replace(/[()]/g, '') || '',
            url: link?.href || ''
        };
    });
}''')
import json
result = json.dumps(results, indent=2)
"
brow session delete 1

Result

Because the google profile persists your login, you get personalized results -no cookie banners, no sign-in walls, just data.

Commands

Daemon

brow daemon start [--port 19987]
brow daemon stop
brow daemon status

Sessions

brow session new [--profile <name>] [--headed]
brow session list
brow session delete <id>

Navigation

brow -s <id> navigate <url>
brow -s <id> wait <selector>
brow -s <id> wait --load

Observation

brow -s <id> snapshot [--search <regex>] [--locator <selector>]
brow -s <id> screenshot [--full] [--path <file>]
brow -s <id> html [--locator <selector>] [--search <regex>]
brow -s <id> logs [--search <regex>] [--count <n>]
brow -s <id> url

Interaction

brow -s <id> click <selector>
brow -s <id> fill <selector> <value>
brow -s <id> type <text>
brow -s <id> key <key>            # Enter, Tab, Meta+a
brow -s <id> hover <selector>
brow -s <id> scroll <pixels>
brow -s <id> scroll-to <selector>
brow -s <id> drag <from> <to>
brow -s <id> upload <selector> <filepath>

Pages

brow -s <id> page list
brow -s <id> page new [url]
brow -s <id> page close [index]
brow -s <id> page switch <index>

Profiles & State

brow profile list
brow profile delete <name>
brow state save <name> -s <id>
brow state restore <name> -s <id>
brow state list

Eval & Run

brow eval -s <id> <code>              # inline, one-off
brow run -s <id> workflow.py --arg k=v  # reusable, from a file

Variables available: page, context, browser, state, pages, plus args (from --arg) for run.

Selectors

Playwright selector syntax:

• CSS: button.submit, #login
• Text: text=Login
• Role: role=button[name="Save"]
• XPath: xpath=//div

Architecture

  ┌─────────────────────────────────────────────────────────────────┐
  │  Agent (Claude Code, script, etc.)                              │
  │                                                                 │
  │  brow session new --headed          ← start browser             │
  │  brow navigate -s 1 "https://..."   ← go to page               │
  │  brow snapshot -s 1                 ← read page (a11y tree)     │
  │  brow click -s 1 "text=Login"       ← interact                  │
  │  brow fill -s 1 "#email" "me@..."   ← fill form                 │
  │  brow screenshot -s 1               ← capture screen            │
  │  brow eval -s 1 "await page..."     ← escape hatch              │
  │  brow session delete 1              ← cleanup                   │
  └──────────────┬──────────────────────────────────────────────────┘
                 │ HTTP (localhost:19987)
                 ▼
  ┌──────────────────────────────────────┐
  │  brow daemon (FastAPI + uvicorn)     │
  │                                      │
  │  ┌──────────┐  ┌──────────────────┐  │
  │  │ Session 1 │  │ ProfileManager   │  │
  │  │ (browser) │  │ ~/.brow/profiles │  │
  │  ├──────────┤  └──────────────────┘  │
  │  │ Session 2 │                       │
  │  │ (browser) │  ┌──────────────────┐  │
  │  └──────────┘  │ StateManager     │  │
  │                 │ ~/.brow/states   │  │
  │                 └──────────────────┘  │
  └──────────────┬───────────────────────┘
                 │ CDP (Chrome DevTools Protocol)
                 ▼
  ┌──────────────────────────────────────┐
  │  Chromium (via Playwright)           │
  │                                      │
  │  ┌────────┐ ┌────────┐ ┌────────┐   │
  │  │ Page 1 │ │ Page 2 │ │ Page 3 │   │
  │  └────────┘ └────────┘ └────────┘   │
  └──────────────────────────────────────┘

• Daemon auto-starts on first brow command
• Persistent Chromium profiles for login session survival
• One browser per session, full isolation
• Headless by default, --headed to watch

Configuration

Benchmarks

16 fixture tasks, Claude Sonnet via AWS Bedrock, 1 run each. Compared against playwright-cli, MCP Playwright, agent-browser, and browser-use.

brow leads on success rate across both suites. On token efficiency, brow leads the 16-task fixture suite (68K avg); agent-browser leads across all 22 tasks (69K avg) — brow's 22-task average is inflated by a single live task where the agent didn't use snapshot filtering. agent-browser wins on wall-clock speed. browser-use runs its own agent loop — included for completeness.

Full results and per-task breakdown →

Resource Usage

~150-300MB per Chromium instance. 10 sessions = ~2-3GB.

License

Elastic License 2.0 (releases <=1.2.0 remain MIT-licensed).