airbrow

Ultra-fast browser CLI for LLM agents. Rust/CDP based.

AI Coding Assistants

Add the skill to your AI coding assistant for richer context:

npx add-skill RenKoya1/airbrow

Works with Claude Code, Codex, Cursor, Gemini CLI, GitHub Copilot, Goose, OpenCode, and Windsurf.

Benchmark Results (2026-01-28)

| Metric | airbrow | agent-browser | playwright-cli | |--------|------|---------------|----------------| | Memory | 2.3MB | 1.9MB* | 60MB | | Context | ~6KB | ~90KB | ~89KB | | Binary | 1.8MB | 670KB* | Node.js |

*agent-browser CLI only; requires Playwright backend

Context Size Comparison (Real Data)

| Tool | Hacker News | GitHub | Total | vs airbrow | |------|-------------|--------|-------|---------| | airbrow | ~3KB | ~3KB | ~6.6KB | 1x | | agent-browser | ~60KB | ~30KB | ~90KB | 14x larger | | playwright-cli | ~58KB | ~31KB | ~89KB | 13x larger |

14x smaller context. 25x less memory. Zero dependencies.

Install

From source

cargo build --release
./target/release/airbrow

Via npm

npm install -g airbrow
airbrow install  # Download Chromium (optional if Chrome installed)
airbrow

Quick Start

> open                       # Start browser
> goto https://example.com   # Navigate
> observe                    # Get interactive elements
{"e":[["r1","a","More info"],["r2","btn","Search"]],"n":2,"t":"Example","u":"https://example.com/"}
> click r1                   # Click by ref
> fill r2 "hello world"      # Fill input
> screenshot shot.png        # Take screenshot
> quit                       # Close

Comparison with Competitors

vs agent-browser (Vercel)

| Feature | airbrow | agent-browser | |---------|------|---------------| | Focus | Speed & minimal context | Semantic understanding | | Context size | ~6KB | ~90KB (14x larger) | | Element refs | r1, r2 | @e1, @e2 | | Output | Compact JSON | Full YAML tree | | PII redaction | Manual | Automatic | | Use case | LLM agents | Complex semantic tasks |

vs playwright-cli (Microsoft)

| Feature | airbrow | playwright-cli | |---------|------|----------------| | Focus | LLM optimization | Full Playwright API | | Memory | 2.3MB | 60MB (25x more) | | Context | ~6KB | ~89KB (13x larger) | | Startup | 1.8s | 5.6s (3x slower) | | Tracing | No | Yes | | MCP support | No | Yes | | Use case | LLM agents | MCP/IDE integration |

Summary

| Metric | airbrow | Competitors | Advantage | |--------|------|-------------|-----------| | Context | ~6KB | ~90KB | 14x smaller | | Memory | 2.3MB | 60MB | 25x less | | Dependencies | None | Node.js+Playwright | Zero |

Commands (50+)

Session

| Command | Description | |---------|-------------| | open | Start browser (add --headed to show) | | reconnect | Reconnect to existing browser | | quit | Close browser and exit |

Navigation

| Command | Description | |---------|-------------| | goto <url> | Navigate to URL | | back | Go back | | forward | Go forward | | reload | Reload page |

Observation

| Command | Description | |---------|-------------| | observe | Get interactive elements with refs (r1, r2...) | | snap | Get URL and title only | | snapshot | Full page with text content | | content | Get main HTML |

Interaction

| Command | Description | |---------|-------------| | click <ref> | Click element | | fill <ref> <text> | Fill input field | | type <ref> <text> | Type character by character | | press <key> | Press key (Enter, Tab, Escape...) | | hover <ref> | Hover over element |

Scroll

| Command | Description | |---------|-------------| | scroll [up\|down\|<px>] | Scroll viewport | | scrollto <ref> | Scroll element into view | | top | Scroll to top | | bottom | Scroll to bottom | | pageup | Page up | | pagedown | Page down |

Tabs

| Command | Description | |---------|-------------| | newtab [url] | Open new tab | | tabs | List all tabs | | tab <n> | Switch to tab | | closetab [n] | Close tab |

Network

| Command | Description | |---------|-------------| | block <pattern> | Block URLs matching pattern | | unblock | Clear all blocks | | blockimages | Block image loading | | blockmedia | Block video/audio | | blockfonts | Block web fonts |

Cookies & Storage

| Command | Description | |---------|-------------| | cookies [domain] | Get cookies | | setcookie <n> <v> <d> | Set cookie | | clearcookies | Clear all cookies | | storage get <key> | Get localStorage item | | storage set <k> <v> | Set localStorage item | | storage clear | Clear localStorage |

Output

| Command | Description | |---------|-------------| | screenshot [path] | Save PNG screenshot | | pdf [path] | Save PDF | | eval <js> | Execute JavaScript |

Viewport & Media

| Command | Description | |---------|-------------| | viewport <w> <h> | Set viewport size | | mobile | Mobile emulation (375x667) | | dark | Dark color scheme | | light | Light color scheme |

Wait

| Command | Description | |---------|-------------| | wait <ms> | Wait milliseconds | | wait <ref> [timeout] | Wait for element |

Selectors

| Type | Example | Description | |------|---------|-------------| | Ref | r1, r2 | From observe output | | CSS | #id, .class | CSS selector | | XPath | //div[@id='x'] | XPath expression | | Text | text:Submit | By text content | | Role | role:button | By ARIA role | | Label | label:Email | By label text | | Placeholder | placeholder:Search | By placeholder |

Output Format

Minimal JSON optimized for LLM context:

{
  "e": [["r1","btn","Submit"],["r2","in","Email"]],
  "n": 2,
  "t": "Page Title",
  "u": "https://example.com"
}

• e: Elements array [ref, type, label]
• n: Element count
• t: Page title
• u: Current URL

Element types: btn (button), in (input), a (link), sel (select), txt (text)

Output Comparison

airbrow (Compact JSON):

{"e":[["r1","a","Hacker News"],["r2","a","new"]],"n":50,"t":"Hacker News","u":"https://news.ycombinator.com/"}

agent-browser (Full YAML Tree):

- document:
  - table:
    - rowgroup:
      - row "Hacker News":
        - cell "Hacker News" [ref=e1]:
          - link "new" [ref=e5]:
            - /url: newest

Result: airbrow output is 14x smaller

LLM Agent Integration

Minimal context example

# Python pseudocode
def browse(cmd):
    return subprocess.run(["airbrow"], input=cmd, capture_output=True).stdout

browse("open")
browse("goto https://example.com")
result = json.loads(browse("observe"))
# result: {"e":[["r1","btn","Login"]],"n":1,"t":"Example","u":"..."}
browse("click r1")

Why minimal context matters

| Tool | observe output | Tokens (~) | |------|----------------|------------| | airbrow | ~200 chars | ~50 | | agent-browser | ~3000 chars | ~750 | | playwright-cli | ~3000 chars | ~750 |

Smaller context = Lower cost + Faster response + Better focus

Feature Comparison

| Feature | airbrow | agent-browser | playwright-cli | |---------|:----:|:-------------:|:--------------:| | Navigation | ✅ | ✅ | ✅ | | Click/Input | ✅ | ✅ | ✅ | | Screenshot | ✅ | ✅ | ✅ | | PDF Output | ✅ | ✅ | ✅ | | Tab Management | ✅ | ✅ | ✅ | | Network Blocking | ✅ | ✅ | ✅ | | Cookies/Storage | ✅ | ✅ | ✅ | | Mobile Emulation | ✅ | ✅ | ✅ | | Dark Mode | ✅ | ✅ | ✅ | | Tracing | ❌ | ✅ | ✅ | | Video Recording | ❌ | ✅ | ❌ | | PII Redaction | ❌ | ✅ | ❌ | | MCP Support | ❌ | ❌ | ✅ |

Security

| Aspect | airbrow | agent-browser | playwright-cli | |--------|------|---------------|----------------| | Sandbox | Chrome default | Playwright | Playwright | | Dependencies | Minimal | Playwright | Node.js + Playwright | | Attack Surface | Low | Medium | High | | PII Protection | Manual | Automatic | Manual |

See Also

• SKILL.md - Full command reference for LLM agents
• ../README.md - Project overview
• ../COMPARISON.md - Detailed benchmark results

License

MIT