WorkBuddy 5.8.4 unifies three connector paths around MCP
A technical analysis of WorkBuddy 5.8.4 traces how its desktop capability bus, third-party connector packages, and user-written MCP servers are exposed to models through a unified connector interface. It says the paths retain separate credential storage and approval mechanisms, while describing controls such as loopback validation, route-specific tokens, context signatures, session allowlists, approval hooks, and configuration-hash-based consent. The author notes that the assessment is based on mechanism analysis rather than real OAuth testing, token theft, or local port scanning.
The article identifies where WorkBuddy simplifies the model-facing tool interface while preserving three distinct trust and credential boundaries internally, a design issue relevant to developers integrating AI agents with local and third-party services.
用 WorkBuddy 时你可能做过三件事:让它读 Gmail 邮箱、把一个站点部署到云端、或者自己在 mcp.json 里填一个本地服务。界面上这三件事都叫连接器,模型看到的也都长成 mcp__server__tool。实现上它们走三条完全不同的路:桌面端内置的能力总线、第三方连接器包、你自己写的 MCP server。
WorkBuddy 的做法是统一命名,但凭据怎么存、调用怎么审批,三层各有一套。
这篇按一个连接器在系统里的生命周期走:注册进来 → 拿到调用资格 → 暴露成工具 → 一次调用经过的检查 → 撤销与失效。每个环节都看三层各自的答案。
注册:连接器以什么形式存在#
三条路的进程来源各不相同:内置能力总线跑在桌面主进程里,第三方连接器包自带一个 MCP server,用户自定义 MCP 由用户或插件启动。
第一层不用安装。桌面主进程起一个本地 MCP 主机,服务名 connector-proxy,把平台能力注册成 MCP 工具,路由挂在 /mcp:云服务连接、站点部署、插件市场、技能安装、会话搜索都在上面,比如 connect_cloud_service、workbuddy_sites_deploy、search_and_install_skills、conversation_search、search_plugins。同样用本地主机挂出来的还有 miora(设计、图像、视频)、sheetagent(表格)、weixinpay、ardot、netdrive。