它解决什么问题？
Cline 能读取和修改代码、运行终端命令、使用浏览器和 MCP 工具，并支持 IDE、交互式或无头 CLI、SDK 和多 Agent 看板。模型既可以使用 Cline 服务，也可以自带 API Key 或连接其他提供商。
Plan/Act 把规划和执行分开；默认的人机确认让用户逐次批准文件修改和命令，也可以按类别开启 Auto Approve。Checkpoint 使用独立的影子 Git 仓库保存文件快照，方便比较和恢复。
为什么是 84 分？
What does it solve?
Cline reads and edits code, runs terminal commands, uses a browser and MCP tools, and ships across IDE, interactive or headless CLI, SDK and multi-agent boards. Users can choose Cline services, bring API keys or connect other providers.
Plan/Act separates planning from execution. Human approval can gate file edits and commands, or Auto Approve can be enabled by category. Checkpoints use a separate shadow Git repository for file snapshots, comparisons and recovery.
Why 84?
| 维度 / Dimension | 得分 / Score | 当前依据 / Evidence |
|---|---|---|
| 实用价值 / Utility | 24/25 | IDE、CLI、SDK、浏览器、MCP 与自动化入口覆盖完整 |
| 可用程度 / Usability | 18/20 | 安装入口多，Plan/Act、审批和 Checkpoint 降低上手门槛 |
| 项目健康 / Health | 15/15 | 高频开发、公开文档与安全响应机制完整 |
| 热度动量 / Momentum | 14/15 | 社区规模大，采集当天仍有代码推送 |
| 安全隐私 / Safety | 5/15 | 本地命令、浏览器与 MCP 权限强；官方已有多项安全公告 |
| 成本 / Cost | 8/10 | 核心 Apache-2.0；模型与订阅成本取决于提供商 |
Auto Approve 应该怎么开？
- 先只自动批准读取：官方建议从项目文件读取开始，写入、命令、浏览器和 MCP 按需要逐项开放。
- YOLO 不是快捷设置：它会自动批准任意位置的文件操作、终端命令、浏览器和 MCP；官方文档直接标为危险。
- Checkpoint 不是安全边界：它能恢复项目文件和会话，但不能撤销已经发送的数据、系统设置修改或外部服务操作。
- MCP 要逐个信任：本地服务器可以启动命令，远程服务器可访问外部数据；只安装可信来源并审查调用。
近期安全信号怎么看？
官方 GitHub Security 页面当前列出三项公告，包括 Dashboard 与 Kanban 的 WebSocket 问题，以及一次被篡改的 npm CLI 发布。项目公开披露并持续修复是正面信号，但也说明桌面、CLI、Hub 和插件供应链扩大了攻击面。只使用最新小版本，不在不可信仓库中开启 YOLO。
试用时重点看什么？
用同一个真实仓库分别完成解释、修改、测试和失败恢复，记录批准次数、错误改动、Checkpoint 恢复效果、模型成本与 MCP 调用。高权限自动化应放在一次性容器或受限虚拟机中。
来源与边界
功能与安装来自 Cline 官方仓库；审批、Checkpoint 与 MCP 结论来自官方文档；安全事件来自官方 Security 页面。指标来自 GitHub REST API。尚未完成统一任务实测，初评置信度 78%。
How should Auto Approve be configured?
- Start with reads only: official guidance begins with project-file reads, then opens writes, commands, browser and MCP only as needed.
- YOLO is not a convenience toggle: it approves file operations anywhere, terminal commands, browser actions and MCP. The official documentation labels it dangerous.
- Checkpoints are not a security boundary: they restore project files and task state, not disclosed data, system changes or external side effects.
- Trust MCP servers individually: local servers can launch commands and remote servers reach external data. Install trusted sources and review calls.
How should recent security signals be read?
The official GitHub Security page currently lists three advisories, including WebSocket issues in Dashboard and Kanban plus a modified npm CLI release. Public disclosure and active patching are positive, but the incidents show how desktop, CLI, Hub and plugin surfaces expand the attack area. Stay on the latest minor release and avoid YOLO in untrusted repositories.
What should a trial measure?
Run the same explain-edit-test-recover task in a real repository. Track approvals, incorrect edits, checkpoint recovery, model cost and MCP calls. Put high-permission automation inside a disposable container or restricted VM.
Sources and limits
Capabilities and installation come from the official repository. Approval, checkpoint and MCP behavior comes from the official documentation; incidents come from the official Security page. Metrics come from the GitHub REST API. No standardized task benchmark yet; confidence is 78%.