Skip to main content
Glama
Sign In
workbuddy-subagent-bridge

workbuddy-subagent-bridge

by BenjaminNH
OverviewSchemaSourceRelated ServersDiscussionsAdmin
TypeScript
Local

WorkBuddy Subagent Bridge

WorkBuddy Subagent Bridge 是一个本地 MCP 桥接器，让主 Agent 通过 MCP 调用 WorkBuddy 执行任务。

主 Agent 负责理解目标、拆解任务、审批权限和验收结果；WorkBuddy 负责具体执行。任务完成后，主 Agent 可以读取报告，并继续向同一会话发送修改意见。

本项目是独立的社区工具，与 WorkBuddy 官方没有隶属关系。

适用场景

  • 让主 Agent 把编码、文件修改或诊断任务交给 WorkBuddy。

  • 管理多个独立的 WorkBuddy 任务，并分别查询状态和报告。

  • 在 WorkBuddy 请求修改文件或执行操作时，由主 Agent 审批权限。

  • 让同一 WorkBuddy 会话持续接收反馈，完成“分派—执行—验收—修订”流程。

当前公开版本为 0.1.0，已在 Windows 上完成真实验证。

Related MCP server: workbuddy-mcp

安装

前置条件

  • Windows

  • Node.js 20 或更高版本

  • 已安装并登录 WorkBuddy / CodeBuddy

当前验证环境：WorkBuddy 5.5.6、CodeBuddy CLI 2.137.1、Node.js 22.22.3。

交由 Agent 安装

复制下面的 Prompt 发送给 Agent 即可：

安装并配置 WorkBuddy Subagent Bridge。

项目仓库：https://github.com/BenjaminNH/workbuddy-subagent-bridge
先阅读并执行仓库中的 docs/agent-install.md，完成后返回安装和验证结果。

从 npm 安装

npm install -g workbuddy-subagent-bridge
workbuddy-subagent doctor

doctor 会检查 Node.js、npm、WorkBuddy、CodeBuddy CLI、ACP 连接和模型目录。开始使用前，请确认关键检查项均为 ok。

从本地源码安装

npm install
npm run build
npm pack
npm install -g .\workbuddy-subagent-bridge-0.1.0.tgz
workbuddy-subagent doctor

配置 MCP

生成 Codex 配置文件：

workbuddy-subagent init --client codex --output .\workbuddy-mcp.json

init 只生成配置文件，不会自动修改 Codex 的全局配置。请将生成的 workbuddy-subagent-bridge 条目加入 Codex 的 MCP 配置，然后重启 Codex。

如果目标文件已存在，init 默认拒绝覆盖；确认需要替换时，显式添加 --force。

基本使用

配置完成后，主 Agent 可以创建任务：

请分析当前项目，实现指定功能，并在完成后报告修改内容和测试结果。

典型流程：

  1. 使用 workbuddy_session_start 创建任务。

  2. 使用 workbuddy_session_status 查询任务状态和待处理权限请求。

  3. 使用 workbuddy_session_report 读取最终报告。

  4. 使用 workbuddy_session_send 向同一会话发送反馈或修改要求。

  5. 完成后使用 workbuddy_session_close 关闭本地会话进程。

一个 MCP Bridge 可以管理多个独立任务。每个任务都有自己的 taskId、WorkBuddy session、工作目录、权限模式和报告。

权限模式

权限模式在创建任务时指定，默认是 plan。主 Agent 应根据任务范围选择模式。

模式

说明

plan

只读分析和规划，默认模式

default

需要执行受限操作时逐项请求权限

acceptEdits

自动接受文件编辑，其他操作仍受限制

auto

由 WorkBuddy 自动判断风险

dontAsk

无法交互时拒绝需要询问的操作

delegate

由主 Agent 管理权限请求

bypassPermissions

跳过权限提示，风险较高

fullAccess

跳过全部权限检查，包括危险命令，风险极高

bypassPermissions 和 fullAccess 不会被默认使用，也不会被后续 session_send 自动提升。主 Agent 应在使用前确认工作目录和任务范围。

MCP 工具

工具

用途

workbuddy_session_start

创建 WorkBuddy 任务

workbuddy_session_send

向同一会话发送后续反馈

workbuddy_session_status

查询任务状态和权限请求

workbuddy_session_report

读取当前任务的最终报告

workbuddy_session_permission_reply

回复权限请求

workbuddy_session_cancel

取消任务

workbuddy_session_close

关闭本地会话进程

workbuddy_status

列出任务及其状态

workbuddy_models

查询模型目录、来源和更新时间

workbuddy_doctor

检查本机运行环境

安全边界和当前限制

  • 工作目录必须由调用方明确提供，Bridge 不自动创建 worktree 或合并 diff。

  • 默认使用 plan，高风险权限必须在任务级别显式指定。

  • 日志只记录基础生命周期信息，不记录 prompt、结果正文、文件内容或凭据。

  • ACP 是多轮任务的主路径；CLI 只用于一次性任务或 ACP 启动前失败的兼容场景。

  • ACP prompt 已提交但结果不确定时，不会自动切换 CLI 重放。

  • completed 和 cancelled 是任务状态标签；如果 WorkBuddy 仍支持加载 session，可以继续发送反馈。

  • 报告保存在当前 Bridge 进程内。Bridge 进程重启后的无缝任务恢复不属于当前 MVP 的保证范围。

  • 当前版本不提供远程公网 Agent 服务、A2A、多账号池、多供应商调度、自动 worktree、GUI 自动化或 OpenAI/Anthropic API 兼容层。

开发与验证

npm install
npm test
npm run typecheck
npm run build
npm run doctor

问题反馈

提交问题时，请附带：

  • WorkBuddy 和 CodeBuddy CLI 版本

  • workbuddy-subagent doctor 输出

  • 最小复现步骤

  • 使用的权限模式

  • 相关任务状态和原始错误字符串

许可证

本项目采用 MIT License。

Available Tools

10 tools
workbuddy_doctorB

Check Node, CodeBuddy CLI, and ACP stdio initialization.

Inputs: no named parameters.

workbuddy_modelsB

List models returned by the current WorkBuddy ACP session/new response.

Inputs: no named parameters.

workbuddy_session_cancelC

Cancel the active WorkBuddy session.

Inputs: taskId (required).

workbuddy_session_closeB

Close the local ACP process and mark the task completed.

Inputs: taskId (required).

workbuddy_session_permission_replyA

Reply to a pending ACP permission request. Inspect workbuddy_session_status first and pass the raw response payload expected by the current WorkBuddy version.

Inputs: taskId (required), response (required), requestId (required).

workbuddy_session_reportA

Read the in-memory report produced by a completed prompt. Reports are not persisted in the task registry.

Inputs: taskId (required).

workbuddy_session_sendC

Send a follow-up message to the same WorkBuddy session.

Inputs: taskId (required), message (required).

workbuddy_session_startA

Start a current-version WorkBuddy session. ACP mode returns a taskId immediately by default; poll status and request the report after completion. Set waitForCompletion=true when a blocking report is preferred.

Inputs: cwd (required), prompt (required), backend, modelId, +3 more.

workbuddy_session_statusA

Read task/session status and any pending ACP permission requests without sending a prompt.

Inputs: taskId (required).

workbuddy_statusA

List persisted WorkBuddy tasks and their current lifecycle status.

Inputs: no named parameters.

Tool Schema Changelog

Recent tool additions, removals, and schema changes observed during successful MCP inspections.

Loading tool schema changelog…

    Deploy Server
    One endpoint for every MCP toolRayrun runs every MCP server behind one URL and holds the credentials. Each call is decided per client, tool, and argument, then recorded.ray.runAd

    TDQS

    A3.5/5.0

    Scored across 10 tools

    Disambiguation4/5

    Most tools target distinct session actions (start, send, cancel, close, permission reply, report) and are easy to tell apart. The only potential confusion is between workbuddy_session_status (per-session state) and workbuddy_status (persisted task registry), which are related but serve different purposes.

    Naming Consistency4/5

    All tools share the workbuddy_ prefix and mostly follow a predictable pattern: workbuddy_session_<verb> for session operations, plus standalone nouns (status, models, doctor). The mix of verb-based session actions with noun-only utility tools is a minor deviation, but the overall convention is clear.

    Tool Count5/5

    Ten tools is well within the ideal range for a session-bridge server and each one maps to a distinct lifecycle stage or utility concern. No tool feels redundant or unnecessary for the stated purpose.

    Completeness4/5

    The session lifecycle is well covered: start, send, status, cancel, close, permission reply, and report. The persisted task listing, model enumeration, and environment diagnostic round out the surface. Minor gaps exist, such as no explicit session resume or task-triggered report fetch, but agents can generally work around these.

    Maintenance

    ActivitySlowing
    ResponsivenessNo recent community issues

    Resources

    Related MCP Connectors

    Related MCP Servers

    • workbuddy-mcp

      LinHaiJ
      A
      license
      A
      quality
      C
      maintenance
      Enables AI coding agents like Claude Code, Codex, Cursor, and OpenCode to delegate tasks to the WorkBuddy CLI as a sub-agent via an MCP tool, with one-command setup and configurable model/permission options.
      1
      309 npm
      7
      MIT

    • agentmux

      seaweedsoup98
      A
      license
      B
      quality
      B
      maintenance
      Enables any MCP host to delegate work to multiple coding-agent CLIs such as Codex, Claude Code, and Antigravity as subagents, preserving native sessions and supporting team-based supervision and agent-to-agent messaging.
      17
      9 npm
      Apache 2.0

    • Sub-Agents MCP Server

      rorato10720
      A
      license
      Not graded
      quality
      B
      maintenance
      Lets any MCP-compatible client run reusable Markdown-defined coding agents through a chosen backend CLI, returning their results to the same conversation. It enables delegation of coding tasks, shared agent definitions, permission modes, and cross-call session continuation.
      182 npm
      MIT