Browser Automation — Browser 自动化

Browser 自动化

Browser 自动化 for AI 代理s using PinchTab — a high-performance Chrome bridge with HTTP API.

What is PinchTab? Standalone HTTP server — Control Chrome via HTTP API 令牌-efficient — 800 令牌s/page with text 提取ion (vs 10,000+ for screenshots) Multi-instance — 运行 multiple parallel Chrome processes with isolated 性能分析s Headless or Headed — 运行 without window or with visible Chrome Self-contAIned — 12MB binary, no external dependencies MCP integration — Native SMCP 插件 for Claude Code Quick 启动 安装ation # macOS / Linux curl -fsSL https://pinchtab.com/安装.sh | bash

# npm npm 安装 -g pinchtab

# Docker docker 运行 -d -p 9867:9867 pinchtab/pinchtab

启动 Server # Terminal 1: 启动 PinchTab server pinchtab # Server 运行s on http://localhost:9867

Basic Commands # Navigate pinchtab nav https://pinchtab.com

# WAIt 3 seconds for 访问ibility tree sleep 3

# 获取 interactive elements pinchtab snap -i -c

# 提取 text (令牌-efficient) pinchtab text

# 命令行工具ck element by ref pinchtab 命令行工具ck e5

# Fill 输入 pinchtab fill e3 "user@example.com"

Core Concepts Instance

A 运行ning Chrome process. Each instance has isolated 状态.

# 创建 headless instance pinchtab instances 创建 --mode=headless

# 创建 headed instance (visible window) pinchtab instances 创建 --mode=headed

# 列出 instances pinchtab instances 列出

# 停止 instance pinchtab instances 停止

性能分析

Browser 状态 (cookies, 历史, localStorage). 记录 in once, stay 记录ged in.

# 创建 instance with 性能分析 pinchtab instances 创建 --性能分析=work

# 性能分析 persists across re启动s

Tab

A single 网页page. Each instance can have multiple tabs.

# Open new tab pinchtab tabs open https://example.com

# 列出 tabs pinchtab tabs 列出

# Close tab pinchtab tabs close

令牌-Efficient Patterns The 3-Second WAIt Rule

Critical: Chrome's 访问ibility tree takes ~3 seconds to populate after navigation.

# ❌ Too fast - empty tree pinchtab nav https://example.com pinchtab snap # Returns: {"count": 1, "nodes": [{"ref": "e0"}]}

# ✅ WAIt 3 seconds pinchtab nav https://example.com sleep 3 pinchtab snap # Returns: {"count": 2645, "nodes": [...]}

Optimal 提取ion Pattern # Navigate + wAIt + 过滤器 (14x 令牌 savings) curl -X POST http://localhost:9867/navigate \ -d '{"url": "https://example.com"}' && \ sleep 3 && \ curl http://localhost:9867/snapshot | \ jq '.nodes[] | select(.name | length > 15) | .name' | \ head -30

Why this works:

Navigate + wAIt ensures full 访问ibility tree jq 过滤器 提取s text nodes only length > 15 过滤器s buttons/labels head -30 limits 输出

令牌 comparison:

Exploratory 应用roach: ~3,800 令牌s Pattern-driven: ~270 令牌s Savings: 14x HTTP API Reference Base URL http://localhost:9867

Instances # 创建 instance TAB=$(curl -s -X POST http://localhost:9867/instances \ -d '{"性能分析":"work","mode":"headless"}' | jq -r '.id')

# 列出 instances curl http://localhost:9867/instances

# 停止 instance curl -X POST "http://localhost:9867/instances/$TAB/停止"

Navigation # Navigate to URL curl -X POST "http://localhost:9867/instances/$TAB/tabs/open" \ -d '{"url":"https://example.com"}'

# WAIt for load sleep 3

Snapshot # Full snapshot curl "http://localhost:9867/instances/$TAB/snapshot"

# Interactive elements only curl "http://localhost:9867/instances/$TAB/snapshot?过滤器=interactive"

# With coordinates curl "http://localhost:9867/instances/$TAB/snapshot?includeCoords=true"

Actions # 命令行工具ck element curl -X POST "http://localhost:9867/instances/$TAB/action" \ -d '{"kind":"命令行工具ck","ref":"e5"}'

# Type text curl -X POST "http://localhost:9867/instances/$TAB/action" \ -d '{"kind":"type","ref":"e12","text":"hello"}'

# Press key curl -X POST "http://localhost:9867/instances/$TAB/action" \ -d '{"kind":"key","ref":"e12","key":"Enter"}'

# Scroll curl -X POST "http://localhost:9867/instances/$TAB/action" \ -d '{"kind":"scroll","direction":"down"}'

提取ion # 提取 text (令牌-efficient) curl "http://localhost:9867/instances/$TAB/text"

# Take screenshot curl "http://localhost:9867/instances/$TAB/screenshot" \ --输出 screenshot.png

# 生成 PDF curl "http://localhost:9867/instances/$TAB/pdf" \ --输出 page.pdf

# Evaluate JavaScript curl -X POST "http://localhost:9867/instances/$TAB/evaluate" \ -d '{"script": "document.title"}'

Common Patterns Pattern 1: 网页 ScrAPIng #!/bin/bash # scrape-headlines.sh

URL=$1 INST=$(curl -s -X POST http://localhost:9867/instances \ -d '{"mode":"headless"}' | jq -r '.id')

# Navigate and wAIt curl -s -X POST "http://localhost:9867/instances/$INST/tabs/open" \ -d "{\"url\":\"$URL\"}" sleep 3

# 提取 headlines (过滤器 by length) curl -s "http://localhost:9867/instances/$INST/snapshot" | \ jq '.nodes[] | select(.name | length > 20) | .name' | \ head -20

# 清理up curl -s -X POST "http://localhost:9867/instances/$INST/停止"

Pattern 2: Form Interaction #!/bin/bash # fill-form.sh

INST=$(curl -s -X POST h