Claude Computer Use

Anthropic 官方桌面自动化 API（beta）。通过截屏 + Claude 视觉理解 + 坐标操作控制任意桌面应用。

何时使用

这是终极后备方案，仅在以下情况使用：

1. peekaboo (macOS) / pywinauto (Windows) / browser (网页) 无法操作目标应用
2. 需要视觉判断（看屏幕内容决定下一步）
3. 目标应用不暴露 Accessibility API
4. 用户明确要求精确坐标控制

优先使用轻量方案（成本更低、速度更快）：

• 网页 → browser tool
• macOS 桌面 → peekaboo
• Windows 桌面 → pywinauto

支持的操作

工作流程

循环:
  1. 截屏 (screencapture / pyautogui.screenshot)
  2. 发送截图给 Claude Computer Use API
  3. Claude 分析截图，返回 tool_use (action + coordinates)
  4. 在本地执行操作 (cliclick/xdotool/pyautogui)
  5. 如果任务未完成，回到步骤 1

完整示例

macOS 实现

import anthropic
import subprocess
import base64
import json

client = anthropic.Anthropic()

def screenshot():
    """截取 macOS 屏幕"""
    subprocess.run(["screencapture", "-x", "/tmp/screen.png"], check=True)
    with open("/tmp/screen.png", "rb") as f:
        return base64.b64encode(f.read()).decode()

def execute_action(action_type, **kwargs):
    """执行鼠标/键盘操作 (macOS 使用 cliclick 或 AppleScript)"""
    if action_type == "left_click":
        x, y = kwargs["coordinate"]
        subprocess.run(["cliclick", f"c:{x},{y}"])
    elif action_type == "type":
        text = kwargs["text"]
        subprocess.run(["cliclick", f"t:{text}"])
    elif action_type == "key":
        key = kwargs["key"]
        # 通过 AppleScript 发送按键
        subprocess.run(["osascript", "-e",
            f'tell application "System Events" to keystroke "{key}"'])
    elif action_type == "scroll":
        direction = kwargs.get("direction", "down")
        amount = kwargs.get("amount", 3)
        delta = -amount if direction == "down" else amount
        # AppleScript 模拟滚动
        subprocess.run(["osascript", "-e",
            f'tell application "System Events" to scroll area 1 by {delta}'])
    elif action_type == "mouse_move":
        x, y = kwargs["coordinate"]
        subprocess.run(["cliclick", f"m:{x},{y}"])
    elif action_type == "screenshot":
        pass  # 下一轮循环自动截屏

def computer_use_loop(task: str, max_iterations: int = 10):
    """Computer Use 主循环"""
    messages = [{"role": "user", "content": [
        {"type": "text", "text": task}
    ]}]

    for i in range(max_iterations):
        # 1. 截屏
        img_b64 = screenshot()

        # 2. 追加截图到消息
        if i > 0:
            messages.append({"role": "user", "content": [
                {"type": "tool_result", "tool_use_id": tool_use_id,
                 "content": [{"type": "image", "source": {
                     "type": "base64", "media_type": "image/png", "data": img_b64
                 }}]}
            ]})
        else:
            messages[0]["content"].append({
                "type": "image", "source": {
                    "type": "base64", "media_type": "image/png", "data": img_b64
                }
            })

        # 3. 调用 Claude Computer Use API
        response = client.beta.messages.create(
            model="claude-sonnet-4-6",
            max_tokens=1024,
            tools=[{
                "type": "computer_20250124",
                "name": "computer",
                "display_width_px": 1920,
                "display_height_px": 1080,
            }],
            messages=messages,
            betas=["computer-use-2025-01-24"],
        )

        # 4. 解析响应
        messages.append({"role": "assistant", "content": response.content})

        # 检查是否完成
        if response.stop_reason == "end_turn":
            print("任务完成")
            break

        # 5. 执行操作
        for block in response.content:
            if block.type == "tool_use":
                tool_use_id = block.id
                action = block.input.get("action")
                print(f"  执行: {action} {block.input}")
                execute_action(action, **block.input)

    return messages

Windows 实现

import pyautogui
import time

def screenshot_win():
    """截取 Windows 屏幕"""
    img = pyautogui.screenshot()
    img.save("/tmp/screen.png")
    with open("/tmp/screen.png", "rb") as f:
        return base64.b64encode(f.read()).decode()

def execute_action_win(action_type, **kwargs):
    """执行操作 (Windows 使用 pyautogui)"""
    if action_type == "left_click":
        x, y = kwargs["coordinate"]
        pyautogui.click(x, y)
    elif action_type == "right_click":
        x, y = kwargs["coordinate"]
        pyautogui.rightClick(x, y)
    elif action_type == "double_click":
        x, y = kwargs["coordinate"]
        pyautogui.doubleClick(x, y)
    elif action_type == "type":
        pyautogui.typewrite(kwargs["text"], interval=0.02)
    elif action_type == "key":
        pyautogui.hotkey(*kwargs["key"].split("+"))
    elif action_type == "scroll":
        amount = kwargs.get("amount", 3)
        direction = kwargs.get("direction", "down")
        clicks = -amount if direction == "down" else amount
        x, y = kwargs.get("coordinate", (960, 540))
        pyautogui.scroll(clicks, x, y)
    elif action_type == "mouse_move":
        x, y = kwargs["coordinate"]
        pyautogui.moveTo(x, y)
    elif action_type == "left_click_drag":
        sx, sy = kwargs.get("start_coordinate", pyautogui.position())
        x, y = kwargs["coordinate"]
        pyautogui.moveTo(sx, sy)
        pyautogui.drag(x - sx, y - sy, duration=0.5)

API 参数

tools=[{
    "type": "computer_20250124",  # 或 computer_20251124 (Opus 4.6)
    "name": "computer",
    "display_width_px": 1920,     # 屏幕宽度
    "display_height_px": 1080,    # 屏幕高度
    "display_number": 1,          # X11 显示号（可选）
}]
betas=["computer-use-2025-01-24"]  # 或 computer-use-2025-11-24 (Opus 4.6)

成本估算

• 每次截屏约 1000-3000 tokens（取决于分辨率）
• 典型任务（5-10 次截屏循环）：约 10,000-30,000 tokens
• 建议：降低截屏分辨率（1280x800 效果最佳）可显著减少成本

安全规则

• 每次操作前 HITL 确认：告知用户即将在屏幕上执行什么操作
• 不自动输入密码：需要登录时提示用户手动操作
• 不操作金融应用：涉及支付/转账时必须用户确认
• 限制循环次数：max_iterations 防止无限截屏循环
• 最小权限：不要给 Claude 管理员权限