multi-agent-bridge
Orchestrate multiple Agent CLIs (Claude Code, Codex, Qwen, DSH, opencode, Qoder) into a collaborative team via a zero-dependency MCP server with task DAGs, shared memory, a message bus, and observability tooling.
Task orchestration — create/list/claim/complete/fail/supersede/reassign tasks with dependency DAGs, deliverable + acceptance criteria, and human approval gates (
task_approve).Long-task control — interrupt (kill process tree), resume via
session_id, heartbeat, escalate decisions up, decide them down, fork alternative paths, and dynamically rewire dependencies.Workflow automation — auto-decompose goals (
workflow_plan), land staged DAGs (workflow_start), and self-heal viaworkflow_evolve(fork/append/insert/degrade/rollback/branch with guardrails).Shared memory — cross-process KV (
shared_memory_*), append-only handoff notes (shared_notes_*), and vector semantic recall (memory_add/search/list/delete/stats/promote,task_sediment).Message bus — persistent inboxes with FIFO + 60s leases, real-time wake-up (
inbox_wait), broadcast (to="*"), topics, priorities, history replay.Worker dispatch — invoke any registered agent via
run_*/agent_invokewith 429 backoff, model rotation, busy-time failover to idle workers, andfork_on_failfallback.Agent registry & selection —
agent_scandetects installed CLIs,agent_listshows capabilities/strengths,agent_evalproduces 5-dim quality metrics.Quality & arbitration —
run_verifyLLM-as-judge scoring with degrade band,result_arbitratemulti-worker conflict resolution (majority → expert-weighted → LLM arbiter).Observability & safety —
bridge_statsruntime metrics,bridge_checkpointstate snapshots, advisoryfile_lock_*,safe_scanblocklist gating.File & media access — sandboxed
read_file/list_dir/project_search,vision_analyzeimage understanding,dsh_readfor DSH session history.
Enables OpenAI-based worker agents such as Codex to be registered and dispatched, with parallel execution, automatic retries, and fallback routing.
Click on "Deploy Server".
Wait a few minutes for the server to deploy. Once ready, it will show a "Started" state.
In the chat, type
@followed by the MCP server name and your instructions, e.g., "@multi-agent-bridge orchestrate claude and codex to refactor the monorepo and share notes"
That's it! The server will respond to your query, and you can continue using it as needed.
Here is a step-by-step guide with screenshots.
多智能体协作系统
Multi-Agent Bridge
把多个 Agent CLI 编排成一支可协作的团队 —— 零三方依赖的 MCP Server + 多端可视化面板。
任务 DAG 编排 · 共享记忆 · 消息总线 · 多端面板 · 跨 Agent 协作
Related MCP server: ATMcp
这是什么(What)
multi-agent-bridge 是一个多智能体协作系统,核心是一个纯 Node.js 标准库实现的 MCP Server,它把多个相互独立的 Agent CLI —— Claude Code、Codex、Qwen、opencode、DSH、Qoder、Qoder CN —— 接进同一张协作网络。
它们本来各自为政、互不通信。这个系统给它们装上「公共设施」和「可视化指挥中心」:
设施 | 解决的问题 |
🗂️ 任务队列 | 把大目标拆成有依赖关系的任务 DAG,由多个 Agent 认领、接力完成 |
🧠 共享记忆 | KV 存储 + 交接笔记 + 向量语义检索,知识在 Agent 之间真正流动 |
✉️ 消息总线 | 持久化收件箱 + 实时唤醒,Agent 之间可以互相发消息、传信号 |
🖥️ 多端面板 | Web 控制台 / IDE 侧边栏 / Qoder 插件,三套界面共享同一套组件 |
在它之上,你还能得到:429 限流自动绕避、多 worker 并行派发、工作流生命周期管理、长任务暂停/续跑,以及一套让「多个 AI 协作完成一个目标」真正落地的编排原语。
为什么(Why)
多 Agent 协作通常意味着「自己写胶水代码」——拼 shell 脚本、粘数据库、手写轮询、手动处理限流。这套系统把所有这些收拢成一个标准化的 MCP Server + 可视化面板:
零三方依赖:只用 Node.js 内置模块(
child_process/fs/readline…),npm install拉 0 个包,部署即拷文件。统一协议:所有 Agent 通过同一套 MCP 工具交流,而不是 N 套私有格式。
可观测:任务状态、心跳、统计、面板一览无余,协作不再是一团黑箱。
抗压:限流(429)、超时、假成功、进程卡死,这些「长任务现实问题」都被内置机制兜住。
多端一致:Web 面板、DSH 侧边栏、Qoder 插件共享同一套 React 组件库,体验统一。
多端面板生态
🖥️ Web 控制台(React 版)
三栏布局的完整协作控制台,适合桌面端全局监控和深度操作。
左栏:任务列表(进行中 / 已完成 / 待开始 / 已归档 分段式分区)
中栏:任务依赖 DAG 图 + 任务详情面板 + 派发新任务按钮
右栏:工作流列表(卡片式,进度条 + Agent 标签)
六套主题:暗夜 / 白昼 / 护眼绿 / 海洋蓝 / 日落橙 / 典雅紫
实时 SSE 推送 + 轮询兜底
📋 DSH Panel(IDE 侧边栏)
嵌入 DSH 编辑器的侧边栏面板,紧凑高效,编码时随时查看协作状态。
StickyHeader 连接状态 + 全局统计
工作流卡片(分段进度条 + 任务展开 + DAG 内联查看)
归档 / 历史可折叠分区
FAB 悬浮按钮 + 派发任务面板
与 Web 面板共享
bridge-ui组件库,体验一致
🔌 Qoder Panel(Qoder CN 插件)
Qoder 国产 IDE 的面板插件,适配国内开发环境。
基于
bridge-ui内联样式系统,绕过 webview CSP 限制同样支持全部协作功能和六套主题
核心特性(Highlights)
🗂️ 任务队列 + DAG 依赖编排
任务带
deliverable(产物路径)与acceptance_criteria(验收标准),下游 Agent 不用猜「做到哪算完」。支持 依赖图(dependency DAG):任务在全部前置终结前不可被认领。
完整的生命周期原语:
claim / complete / fail / supersede / reassign / fork / depend。自带演进能力:失败自动换备选 Agent(
fork)、遗漏前置反向插入(insert)、已完成段打回重做(rollback)、条件分支(branch),且全部受服务端护栏约束(演进次数上限 + 独立审闸)。
⚡ 多 worker 并行派发 + 备路改派
一个目标可并行派给 ≥2 个视角,产出后经结果仲裁(多数一致优先 → 专家加权 → LLM 仲裁者兜底)收敛出最优解。
agent_invoke支持忙时自动改派到空闲备路 worker,主控永不空转、也不被长任务压垮。竞争式 / 合作式 / 动态路由三种协作范式一键起。
🔁 429 限流自动规避 + 模型轮换 + 指数退避重试
遇到限流 / 超时自动按
Retry-After指数退避重试(默认 2 次、共 3 次尝试)。模型轮换池:单模型持续 429 时自动切换到备用模型(每个 Agent 内置 fallback 列表),绕过单点限流瓶颈。
假成功检测:识别「exit 0 但正文实为上游 5xx」的静默失败,走重试而非误判成功。
🧠 跨 Agent 共享记忆(KV + 笔记 + 向量语义检索)
KV 存储(
shared_memory_*):跨进程共享键值,与独立插件shared-memory共用同一份存储,简单直接。交接笔记(
shared_notes_*):append-only 带时间戳与 tag,天然适合handoff:<id>。向量语义检索(
memory_*):ONNX + sqlite-vec,memory_search按「意思」而非「关键词」跨 Agent 召回;记忆可沉淀(task_sediment)、可提级(memory_promote)。模型按需下载,不进 git。
✉️ 消息总线(持久化收件箱 + 实时唤醒)
消息持久化落盘,FIFO + 60s 租约防并发双处理。
实时唤醒:接收端
inbox_wait挂起时,发送端发消息即被即时唤醒,不靠轮询。支持
topic分组、priority分级、memory自动沉淀、to="*"广播。
🎨 六套主题 + 统一组件库
bridge-ui:共享 React 组件库,DSH Panel / Qoder Panel / Web Panel 三端复用
六套主题:暗夜 Dark / 白昼 Light / 护眼绿 Eye Care / 海洋蓝 Ocean / 日落橙 Sunset / 典雅紫 Elegant
CSS 变量驱动,一键切换,带 anti-flash 防闪烁
🔌 可扩展 Agent Registry(即插即用新 CLI)
开箱即用 7 个 Agent:Claude Code、Codex、Qwen、opencode、DSH、Qoder、Qoder CN,均已内置描述子,零配置自动识别。
agent_scan一键探测:扫描本机已安装的所有 Agent CLI,自动检测可用性、写入注册表;缺失项给出安装引导链接。统一抽象层:所有 Worker 通过统一的
run_<agent>/agent_invoke接口调用,上层编排不关心底层是哪个 CLI。轻松接入新 Agent:新增一个 CLI 只需在
agents-registry.mjs中加一个描述子(启动命令、环境变量映射、限流策略、能力标签),无需改业务代码。能力标签 + 自动选型:每个 Agent 带
capabilities标签和strengths描述,agent_list供主控按需选型,agent_eval出质量/时长/满意度量化指标。自动注册共享记忆:新接入的 Agent 自动获得任务队列、共享记忆、消息总线全部能力,无需逐个对接。
⏱️ 长任务管理(中断 / 续跑 / 状态机)
task_interrupt:人工叫停卡死/跑偏的任务(Windowstaskkill /T /F,UnixSIGTERM→SIGKILL),保留部分输出与session_id。task_resume:复用session_id真·续跑(保留上下文),中断/失败/替代态均可恢复。任务状态机完备:
pending → running → interrupted / escalating / awaiting_approval / completed / failed / superseded。
快速开始(3 步)
要求:Node.js ≥ 18,已安装至少一个 Agent CLI(建议从 Claude Code 开始作主控)。
第 1 步:克隆仓库
git clone https://github.com/songzhifei512/multi-agent-bridge.git
cd multi-agent-bridge第 2 步:配置环境变量
复制模板并按注释填入你的端点与密钥(所有 <占位> 都是示例,替换成你自己的值):
cp config/env.tmpl .env # 或手动抄到 ~/.agents/.env.env 核心项(端点一律用占位符示意,实填你的供应商):
# 主控(必填)
ANTHROPIC_BASE_URL=<PROVIDER_ANTHROPIC_BASE_URL>
ANTHROPIC_AUTH_TOKEN=<YOUR_ANTHROPIC_API_KEY>
BRIDGE_CONTROLLER=claude
# 可选 worker:OpenAI / Qwen / DSH / opencode(按需开启)
# OPENAI_BASE_URL=<PROVIDER_OPENAI_BASE_URL>
# OPENAI_API_KEY=<YOUR_OPENAI_API_KEY>
# QWEN_BASE_URL=<PROVIDER_QWEN_BASE_URL>
# QWEN_API_KEY=<YOUR_QWEN_API_KEY>
# DSH_BASE_URL=<PROVIDER_DSH_BASE_URL>
# DSH_API_KEY=<YOUR_DSH_API_KEY>完整说明见
public-install/ENV_SETUP.md,配置示例见public-install/agents-config-example/.env.example。
第 3 步:挂载到 MCP 客户端
把桥接器注册为你的 Agent CLI 的 MCP Server(模板:config/claude-mcp-config.json.tmpl / config/codex-mcp-config.toml.tmpl)。
以 Claude Code 为例,在配置中加入:
{
"mcpServers": {
"multi-agent-bridge": {
"command": "node",
"args": ["E:/path/to/multi-agent-bridge/bridge/mcp/shared-context-server.mjs"],
"env": {
"BRIDGE_WORK_ROOT": "E:/your/project"
}
}
}
}安装向导也已备好(自动替换路径占位并写入配置):
bash launchers/install.sh # Windows 用 launchers/install-win.bat
挂载后,主控 Agent 就能通过工具看到并调用整套协作能力了。
架构
┌─────────────────────────────────────────────────────────────────┐
│ 你 / 主控 Agent (Claude Code) │
│ 以 MCP 协议调用协作工具 │
└──────────────────────────────┬──────────────────────────────────┘
│ stdio (MCP)
▼
┌─────────────────────────────────────────────────────────────────┐
│ shared-context-server.mjs (MCP Server) │
│ 纯 Node.js 标准库 · 零三方依赖 · 59 个协作工具 │
├───────────────────┬─────────────────┬───────────────────────────┤
│ 🗂️ 任务队列 │ 🧠 共享记忆 │ ✉️ 消息总线 │
│ (DAG 编排) │ KV/笔记/向量 │ (收件箱 + 实时唤醒) │
├───────────────────┴─────────────────┴───────────────────────────┤
│ Worker 派发层(Agent Registry / agent_scan) │
└───────┬───────────┬───────────┬───────────┬───────────┬───────────┬───────────┐
▼ ▼ ▼ ▼ ▼ ▼ ▼
Claude Code Codex Qwen opencode DSH Qoder Qoder CN
推理/架构 批量代码 文档/PPT 备路/并发 多后端执行 全栈编程 国内通义
│
┌──────────┴──────────┐
▼ ▼
┌─────────────────┐ ┌─────────────────┐
│ 🌐 Web 面板 │ │ 📋 DSH 侧边栏 │
│ (React/Vite) │ │ (DSH Plugin) │
└─────────────────┘ └─────────────────┘
│ │
└──────────┬──────────┘
▼
┌──────────────────────┐
│ bridge-ui 组件库 │
│ (12+ 共享组件) │
│ · 6 套主题 │
│ · 双样式系统 │
└──────────────────────┘工具全览(59 个工具 · 5 大类)
🗂️ 任务编排
工具 | 用途 |
| 创建 / 查询任务(支持状态/搜索/排序/分页/批量操作) |
| 认领 / 完成 / 失败生命周期 |
| 标记被替代 / 解锁退回待认领(安全交接) |
| 人工验收门( |
| 中断长任务 / 复用 session 续跑 |
| 分叉子任务 / 动态重算依赖 |
| 决策上浮 / 决策下放 / 里程碑心跳 |
| 完成任务自动沉淀为可检索知识点 |
🧠 共享记忆
工具 | 用途 |
| 跨进程共享 KV 存储(与 |
| append-only 交接笔记(带 tag) |
| 写入 / 向量语义检索记忆 |
| 记忆库管理 |
| 把项目级记忆提级为平台/全局 |
✉️ 消息总线
工具 | 用途 |
| 发消息(含广播 |
| 同步读 / 实时唤醒读(事件驱动,不阻塞服务器) |
| 确认消费(释放 60s 租约) |
| 信号/消息流只读回放 |
🤖 Worker 管理
工具 | 用途 |
| 异步调用各 CLI Agent(自动限流重试 + 模型轮换) |
| 列出注册 Agent / 按名派发(忙时备路改派到空闲 worker) |
| 探测本机已安装的 7 个 worker CLI 并挂载可用性标记 |
| 按 Agent 聚合完成率/质量分/时长/重试/满意度(5 维评估) |
| LLM-as-judge 质量门禁(按标准打分 0–100,支持降级放行) |
| 编排:自动拆解 / 落 DAG / 自适应重规划 |
| 多 worker 冲突结果自动仲裁(多数一致 → 专家加权 → LLM 兜底) |
👁️ 可观测性
工具 | 用途 |
| 运行时可观测(按 Agent 聚合调用/成败/重试/耗时) |
| 状态快照:存 / 列 / 恢复(审计与回滚) |
| 跨进程文件锁(防两 Agent 同修一文件) |
| 共享只读文件访问 |
| 产物安全硬拦(自动审批前把关) |
| 图像理解(场景/文字识别,非条码解码) |
| 只读读取 DSH 历史会话 |
项目结构
multi-agent-bridge/
├── bridge/
│ └── mcp/ # 核心 MCP Server(纯 Node stdlib)
│ ├── shared-context-server.mjs # 主服务入口(59 个工具)
│ ├── bridge-web-panel.mjs # Web 面板 HTTP 服务
│ ├── run-driver.mjs # Worker 派发与执行引擎
│ ├── state-store.mjs # 状态持久化
│ ├── vec_memory.mjs # 向量记忆层(可选)
│ ├── agents-registry.mjs # Agent 注册表
│ └── ...
├── bridge-ui/ # 共享 React 组件库
│ ├── src/
│ │ ├── components/ # StickyHeader / WorkflowCard / TaskRow 等
│ │ ├── hooks/ # useBridgeState / useTheme / useOfflineState
│ │ ├── styles.ts # CSS 样式(class 模式)
│ │ ├── inlineStyles.ts # 内联样式(CSP 兼容模式)
│ │ └── index.ts
│ └── preview/ # 组件预览与截图
├── bridge-web/ # Web 控制台(React + Vite)
│ ├── src/
│ │ ├── components/ # LeftColumn / MiddleColumn / RightColumn
│ │ ├── App.tsx
│ │ └── main.tsx
│ └── dist/index.html # 单文件构建产物
├── dsh-panel/ # DSH 侧边栏插件
│ └── src/client/index.tsx
├── qoder-panel/ # Qoder CN 插件
│ └── src/browser/BridgeConsole.tsx
├── config/ # 配置模板
├── launchers/ # 安装/启动脚本
├── scripts/ # 测试与探测脚本
├── public-install/ # 公开安装文档
├── docs/ # 文档与发布说明
│ ├── guides/ # 指南文档
│ │ ├── troubleshooting.md # 故障排查手册(主手册,13 章全范围覆盖)
│ │ └── dsh-integration-guide.md # DSH 安装与接入指南
│ ├── cookbook/ # 经验食谱
│ │ └── exception-handling.md # 异常处理 Cookbook(设计原则与工程经验)
│ └── releases/ # 版本发布说明
├── skills/ # MCP Skill 定义
├── README.md
├── README_EN.md
├── CHANGELOG.md
├── AGENTS.md
└── package.json什么是 MCP
Model Context Protocol(MCP)是 Anthropic 提出的开放协议,让 AI 助手能以标准方式连接外部数据源和工具。本项目以 MCP Server 形态存在,任何支持 MCP 的 Agent CLI 都可以直接挂载使用。
文档资源
类别 | 文档 | 说明 |
🚀 快速开始 | 公网安装与环境配置指引 | |
🐛 故障排查 | 13 章全范围覆盖:环境/连接/派发/执行/上报/面板/数据/并发/Windows | |
📋 派发排障 | 已并入故障排查手册第 5-6 章 | 派发链路逐级定位 + 各主控特有问题 |
🔧 部署接入 | DSH CLI 安装、MCP Server 接入、dsh-panel 侧边栏安装完整流程 | |
🧑🍳 经验食谱 | 跨项目普适的多 Agent 协作工程经验与设计原则 | |
📝 版本变更 | 各版本功能变更记录 | |
🤝 协作规范 | AI 协作者仓库契约与提交规范 |
许可证
MIT © multi-agent-bridge contributors
Available Tools
63 toolsagent_evalA
Agent 能力评估体系:按 agent 聚合任务记录出 完成率/平均质量分/平均时长/平均重试/满意度 (五等)。只读,供任务路由与选型建议。
Inputs: no named parameters.
agent_invokeA
Invoke any registered agent by name to run a task. Generic path over the Agent Registry — same driver as the run_* tools but name-driven, so new agents (e.g. qwen) need no per-agent tool. Non-blocking. Auto-tracks task. Pass session_id to resume a prior session (the captured id is returned for reuse). auto is honored by agents that support it (codex, qwen); others ignore it. Use agent_list to see available names. Retries automatically on 429/rate-limit/timeout with exponential backoff (default 2 retries, 3 total attempts); set max_retries=0 to disable. 【后台契约】调用可立即返回/被调用方撤回:server 端 Promise 不会因调用方撤回而终止,worker 继续在后台跑到完成,结果落 task.result(及 trace,若 capture_trace);调用方随时可用返回的 task_id 经 task_list 或面板 /api/state 取最终产物,无需阻塞等本次调用返回。给长任务(评估/设计/重构, prompt>2000字)显式传 timeout_sec 600~900 防误杀;传 plan_mode 只读调研不落盘。【默认worker】name 可选:省略 或 指定==控制主控(BRIDGE_CONTROLLER) 时,改从空闲 worker 池轮询派一个(排主控,不压 main;全忙回退主控/或 qwen),返回文案标注实际 worker。【备路】name 明确且该 worker 已忙(agent_live busy/有 running 任务)时自动改派空闲备路 worker(排控制主控,不压 main),返回文案标注改派;same_worker:true 强制精确同名、auto_fallback:false 关备路。
Inputs: prompt (required), auto, name, model, +13 more.
agent_listA
List all registered agents (run_* workers) available to agent_invoke. One agent per line: name + (auto: yes) if it honors the auto param (only codex) + [capabilities] + strengths. Use it to pick the right worker for a task. 不自动派单——只备齐选型数据。
Inputs: no named parameters.
agent_scanA
识别本地已安装的 worker CLI(claude/codex/qwen/opencode/dsh),检测是否可加入 multi-agent,并把可用性挂载到 registry(AGENTS[name].available)。返回每个 agent 的 available + 缺失引导。bin 类型检测 PATH/绝对路径可执行;qwen 属端点类型,检测 QWEN_BASE_URL/OPENAI_BASE_URL 是否配置。available=false 的 worker 被 run_*/agent_invoke 派发时会被拒绝(提示未安装)。
Inputs: refresh.
agent_send_messageA
消息总线(兼容保留): 发一条消息到某 agent 收件箱。持久化落 memory.json mailbox。含【实时唤醒】——若接收端此刻正用 inbox_wait 挂起等待, 立即被唤醒拿到该消息(不等下一轮 poll); 若不在 wait, 消息留存待其下次读。可选 kind/topic/priority/memory(自动沉淀进向量记忆)。返回消息 id。
Inputs: to (required), body (required), from (required), kind, +3 more.
bridge_checkpointA
Save / list / restore named snapshots of the bridge state (memory.json: task queue + file locks + mailbox; 共享 KV/笔记已外置到 ~/.agents/shared-memory/,不随 checkpoint 回滚). For audit trail and rollback before risky bulk operations. Save: pass name (e.g. 'pre-promotion-batch'). List: pass name='list'. Restore (DANGER: overwrites current state): pass name= + restore=true. Snapshots stored as memory.json.checkpoint. alongside memory.json; do not collide with the live file.
Inputs: name (required), restore.
bridge_statsD
Aggregate runtime stats over the task queue (mem.tasks) — per agent (callee): total calls, success (exit=0), failed, timeout, total retries, avg duration ms. Read-only aggregation of data already in memory.json (run_* / agent_invoke write retries + exit_code + timestamps). For diagnosing rate-limit (429) vs sustained overload: a high failure/timeout rate with many retries signals overload, not a task bug. No args.
Inputs: no named parameters.
bus_historyA
消息总线【只读信号/消息流回放】: 列某 agent / topic / kind 最近 N 条(含已 consume), 供面板展示/订阅者审计/历史回放。默认 agent 必填; limit≤200。
Inputs: agent (required), kind, limit, topic.
bus_sendA
消息总线【全能力发送】: to/from/body + kind/topic/priority + memory(自动沉淀向量记忆) + 广播(to="*" 唤醒所有广播订阅者)。含实时唤醒——若接收端正 inbox_wait 挂起则即时命中。返回消息 id。
Inputs: to (required), body (required), from, kind, +3 more.
dsh_readA
Read DeepSeek Harness (DSH) historical sessions. Read-only, delegates to ~/.agents/bin/dsh-read.mjs (zstd-inflate; never mutates DSH data). action='list' lists all sessions (optional keyword filters by title); 'grep ' finds sessions whose title contains keyword; 'get <session_id>' prints the session's condensed [user→assistant→tool] dialogue (pass raw=true for original JSONL). Use when asked to recall/show what a DSH chat session did.
Inputs: action (required), raw, keyword, session_id.
file_lock_acquireA
Acquire an advisory file lock to prevent two agents editing the same file concurrently. Returns ok:true on success, or ok:false with the current holder if already locked. Locks auto-expire (default 30min) so a crashed agent can't hold a lock forever. Path is normalized to absolute.
Inputs: file_path (required), ttl_sec, agent_name.
file_lock_listA
List all currently held (non-expired) file locks. For diagnostics.
Inputs: no named parameters.
file_lock_releaseA
Release a previously acquired file lock. Returns the holder that was released. Advisory lock — cooperative, not enforced.
Inputs: file_path (required).
inbox_ackA
消息总线消费: 把某 agent 的 id(数组)标记 consumed, 释放租约。只有赢得租约的读者可 ack(幂等——ack 已消费/非己有/未过租约期的是 no-op 非错)。消费后不再出现在 inbox_read/inbox_wait。
Inputs: ids (required), agent (required).
inbox_readA
消息总线读(同步, 兼容保留): 列某 agent 未消费(unconsumed)且未被他人租约占用的消息, FIFO 序, 各带 id + 信封(kind/topic/priority)。不消费——读后须 inbox_ack {agent, ids} 确认, 授予/续 60s 租约防并发双处理。实时唤醒用 inbox_wait。
Inputs: agent (required).
inbox_waitA
消息总线【实时唤醒读】(异步阻塞, 事件驱动不阻塞服务器): 读某 agent 消息; 有已 pending → 立即返回(授 60s 租约可 ack); 无 → 挂起等待, 当 bus_send/agent_send_message 发来即被【即时唤醒】返回, 或超时(wait_ms 默认30s 上限120s 返回 {items:[], timed_out:true})。可选 kind/topic 过滤订阅。返回 { items:[{id,from,to,kind,topic,priority,body,created_at}], timed_out }。消费用 inbox_ack。
Inputs: agent (required), kind, topic, wait_ms.
list_dirD
List directory entries (one level). Path confine to BRIDGE_WORK_ROOT per .
Inputs: path, workdir.
memory_addA
Add a memory to the shared vector store with an embedding (paraphrase-multilingual-MiniLM-L12-v2, 384-dim). Stores content + category + source for later semantic recall via memory_search. Category is auto-constructed as layered '::general' from scope/cwd + content platform hints, unless an explicit category is given. Use to sediment cross-agent knowledge (learnings, gotchas, decisions) that should be findable by meaning, not just keywords.
Inputs: content (required), cwd, scope, source, +1 more.
memory_deleteA
Delete memories from the vector store. Mutually exclusive modes (priority id > category > category_prefix): by id (single row), by exact category (all rows in that category), or by category prefix (batch, e.g. 'test:' to clean test entries, or 'project:oldname:' after migration). Returns count deleted + mode. Idempotent — 0 deleted if no match. Passing no args is a no-op (never full-table delete).
Inputs: id, category, category_prefix.
memory_listA
List memories in the vector store (no embeddings returned, only metadata + content preview). Filter by exact category or category prefix (e.g. 'project::' to list one project's memories); paginate with limit/offset. Returns id + category + source + content preview per row. Use memory_stats for counts/breakdown.
Inputs: limit, offset, category, category_prefix.
memory_promoteA
Manually promote a memory to a wider scope (usage-based promotion, manual edition). Changes one memory's category from project:X to platform:Y or global, so future searches find it under the wider scope. Use when an Agent/user judges a project-specific memory is actually cross-project common knowledge. Find the id via memory_list first. Default dry_run=true (preview only, no change) — pass dry_run=false to actually update. to_scope: 'global' | 'platform:' | 'platform:auto' (guess platform from content, fall back to global) | 'project:'. Returns old/new category + whether promoted.
Inputs: id (required), to_scope (required), dry_run.
memory_searchA
Semantic search over the shared vector memory store (paraphrase-multilingual-MiniLM-L12-v2, 384-dim embeddings + sqlite-vec KNN). Returns top-k memories by cosine similarity. By default filters by scope (current project + guessed platform + global) to avoid cross-project noise — pass scope='global' to search only global knowledge, or pass an explicit category for exact-match filtering. Optional min_length filters out short structural segments (e.g. 40). Lazy-loads the ONNX model + sqlite-vec on first call. Complements keyword search (project_search) for 'have I seen something like this before' recall across agents.
Inputs: query (required), cwd, scope, top_k, +2 more.
memory_statsA
Vector memory store stats: total count, per-category breakdown, embedding dim, model path. For diagnostics and before/after bulk operations. Lazy-loads ONNX+sqlite-vec on first call.
Inputs: no named parameters.
project_searchC
Search project with ripgrep (fallback grep). Returns file:line:match.
Inputs: pattern (required), max, glob, path.
read_fileB
Read file content with line numbers. For letting the other agent inspect a file. Path is confined to BRIDGE_WORK_ROOT (or the caller's workdir) per path isolation — out-of-root paths are rejected.
Inputs: file_path (required), limit, offset, workdir.
result_arbitrateA
结果冲突仲裁:多 worker 对同一问题给出不同答案时自动裁决。输入 ≥2 份候选结果,三层裁决:①多数一致优先(结论相同直接过)→ ②专家加权(claude 推理 3 / codex 执行 2 / 其它 1,confidence 可选加权)→ ③仍无胜者时派 LLM 仲裁者(从空闲 worker 池轮询选,排控制主控防自我指涉,全忙回退 qwen)背书。返回 winner + 裁决层 + 各候选权重明细 + 仲裁理由。只裁决不建任务,供 Orchestrator 合并多 worker 并行产出时调用( 配套)。
Inputs: question (required), candidates (required), arbiter, criteria.
run_claudeA
Async call Claude CLI (claude -p headless) to run a task. For reasoning/architecture. Injects auth env. Non-blocking. Auto-tracks task. Pass session_id to resume a prior Claude session (preserves context); the new session id is captured from output and returned for later resumption. Retries automatically on 429/rate-limit/timeout with exponential backoff (default 2 retries, 3 total attempts); set max_retries=0 to disable.
Inputs: prompt (required), model, task_id, workdir, +8 more.
run_codexA
Async call Codex CLI (codex exec) to run a task. For batch codegen/patches. Non-blocking. Auto-tracks task. Pass session_id to resume a prior Codex session (preserves context); the new session id is captured from output and returned for later resumption. Retries automatically on 429/rate-limit/timeout with exponential backoff (default 2 retries, 3 total attempts); set max_retries=0 to disable.
Inputs: prompt (required), auto, model, task_id, +9 more.
run_dshA
Async call DeepSeek DSH CLI (dsh --profile headless) to run a task. Provider-rotatable across the configurable <BACKEND_1>/<BACKEND_2>/<BACKEND_3> backends (same agent loop; pick backend via ~/.dsh settings/patch). Runs with cwd = workdir, which is the workspace-write sandbox boundary: it can REALLY write files and run bash/pwsh inside workdir, non-interactive (no approval stall). Non-blocking. Auto-tracks task. NOTE: no resume (session_id is accepted but starts a fresh run, like opencode). Retries on 429/timeout with exponential backoff (default 2, 3 attempts total); set max_retries=0 to disable. Pass a bounded isolated workdir, never the main repo / a sensitive volume.
Inputs: prompt (required), task_id, workdir, plan_mode, +7 more.
run_qoderB
Async call Qoder CLI (qodercli -p -o json) to run a task. Alibaba Qwen ecosystem full-stack coding agent — for Chinese-optimized software engineering, code review (/review), and /goal long-task mode. Independent rate-limit pool from ByteDance/Anthropic. Real resume via -r (--resume), auto via --permission-mode auto, JSON output. Non-blocking. Auto-tracks task. Pass session_id to resume a prior Qoder session. Retries automatically on 429/rate-limit/timeout with exponential backoff (default 2 retries).
Inputs: prompt (required), auto, model, task_id, +9 more.
run_qoder_cnA
Async call Qoder CN CLI (qoderclicn -p -o json) to run a task. Qoder 国内版(原通义灵码),阿里云通义大模型国内部署 — 合规数据不出境、国内低延迟、中文原生优化。独立于国际版 Qoder 的账号/模型/限流池,国内网络环境下访问更稳定。Real resume via -r, auto via --permission-mode accept_edits + --no-session-persistence, JSON output. Non-blocking. Auto-tracks task. Retries automatically on 429 with exponential backoff.
Inputs: prompt (required), auto, model, task_id, +9 more.
run_qwenA
Async call Qwen CLI (qwen --auth-type openai, Qwen Code) to run a task. For writing/iterating docs & PPT (complements opencode) and image analysis (complements vision_analyze). Uses the qwen 端点 OpenAI-compatible endpoint via QWEN_ENV. --approval-mode auto-edit lets write_file land to disk (auto). Non-blocking. Auto-tracks task. Pass session_id to resume a prior Qwen session (preserves context — qwen has real resume via --resume, unlike opencode); the new session id is captured from JSON output and returned for later resumption. Retries automatically on 429/rate-limit/timeout with exponential backoff (default 2 retries, 3 total attempts); set max_retries=0 to disable.
Inputs: prompt (required), auto, model, task_id, +9 more.
run_verifyA
质量门禁:LLM-as-judge 对产物按 criteria 打分(0-100)。低于 threshold(默认80) 判不通过;但在带宽 degrade_band(默认10) 内(score∈[th-band,th)) 判降级放行(pass + degraded 标记 + verify_degraded 留痕)。打分写回任务 quality_score。默认 evaluator=qwen,可换 claude。可作任务下游 gate。
Inputs: artifact (required), criteria (required), task_id, evaluator, +2 more.
safe_scanA
自动安全硬拦:对一段文本/产物按 blocklist(默认内建危险模式)预扫描。命中即 blocked → 不可自动放行,清单返回命中模式。用于自动审批前的把关; 小模型/人工兜底据此再判断。blocklist 可选传(字符串或正则)。
Inputs: content (required), blocklist.
task_approveA
DAG acceptance gate. Act on a task currently in awaiting_approval: decision=approve (…→completed, releases dependents) or decision=reject (→back to running for redo). Rejects a non-awaiting task. Records approver.
Inputs: task_id (required), decision (required), note, approver, +1 more.
task_claimA
Claim a pending task by id. Transitions pending→running, records claimant. Rejects if not pending (blackboard auto-claim).
Inputs: task_id (required), agent_name.
task_completeA
Mark a running task completed with result. Transitions running→completed. Optional agent_name records who completed (useful if handoff mid-task: A claims, B completes).
Inputs: task_id (required), result, agent_name.
task_createA
Create a new task in the task queue. Returns task_id. Optional deliverable (artifact path) and acceptance_criteria (how to verify done) let the next agent check completion without guessing. Optional dependencies (array of task ids) builds; a dependency DAG: the task stays non-claimable until every dependency is terminal. Explicitly new approval-criteria + require_approval: if require_approval=true, the task is a human-approved approval gate — its completion (task_complete by the implementer) lands in awaiting_approval and stays there, blocking any downstream dependent DAG node, until a human calls task_approve. That's the block-until-ack approval gate.
Inputs: title (required), priority, assigned_to, deliverable, +4 more.
task_decideA
决策下浮()队长/用户在任务 escalating 时下发裁决:写 escalation.decision + decider + decided_at,任务回 running(worker 续跑)。仅对 escalating 态生效;幂等——已决策任务再 decide 拒绝。记录 decider 供审计。
Inputs: choice (required), task_id (required), decider.
task_dependA
前置插入核心原语⚠:动态重算任务依赖(覆盖式修改已有任务的 dependencies)。用于前置插入——fork 出补丁分支 S3' 后,把 S3' 的 dependencies 改为指向新插入的前置 S2.5(set=[原依赖…, S2.5]),实现『反向插入硬性前置 + 重算依赖链』。内置【无环校验】——若 set 引入环(目标任务成为自己的直接/间接依赖)则拒绝;已 terminal(completed/superseded)任务不可改(遵守『不破坏已完成段』演进约束)。不改变状态、不派单,纯依赖图调整。
Inputs: set (required), task_id (required), reason.
task_escalateA
决策上浮()worker 遇方案选择/疑问时向队长/用户上浮:把任务转 escalating 态(运行中阻塞,等人类决策),存 question + options。需持当前 attempt_id(claim 签发;交接/伪造令牌拒绝,防迟到)。escalating 是非终态:监控每周期向用户聚合上报(带 raised_at 年龄),用户回答后由队长 task_decide 下发决策、任务回 running。
Inputs: task_id (required), question (required), options, attempt_id.
task_failC
Mark a task failed with reason (e.g. could not complete).
Inputs: task_id (required), result.
task_forkA
Session 分叉:从 running 任务派生子任务(继承 deliverable/dependencies/acceptance_criteria),原子地把原任务标记 superseded→子任务 id。父任务不 rewrite history。用于探索不同执行路径/决策变更。返回新子任务 id。
Inputs: reason (required), task_id (required), fork_title.
task_heartbeatA
心跳 worker/主控里程碑汇报:给一个 running 任务刷 last_heartbeat_at + heartbeat_n,并把 note 追加进 progress_log(里程碑)。注意:服务端已对每个 run_* 子进程自动心跳(进程活着就跳,无需调用),本工具供能调 MCP 的主控/worker 在生产阶段主动上报里程碑(如『方案已出,等决策』),或手动续活一个进程仍活着但长期无 stdout 的长任务。需持当前 attempt_id(claim 签发;旧/伪造令牌拒绝),防迟到覆盖。
Inputs: task_id (required), note, attempt_id.
task_interruptA
中断一个 running 任务:kill 其 worker 子进程树(Windows 走 taskkill /T /F,Unix 走 SIGTERM→SIGKILL),并把任务置 interrupted 态(保留中断前部分输出 + session_id,供 task_resume 续接)。仅 running 态可中断;终态任务拒绝。适合长任务卡死/跑偏时人工叫停。
Inputs: task_id (required), by.
task_listA
List tasks. Filters: status / search(标题+描述+id 模糊) / task_ids(限定集合,批量查询). 排序: sort=priority(按优先级 high>medium>low)|created(创建时间)|status(状态分组). 分页: limit/offset. 批量操作: batch_action=reassign|fail|supersede 配合 task_ids + batch_agent/batch_reason 一次性对多个任务施效(原子). 不带参数=返回全部(兼容旧行为).
Inputs: sort, limit, offset, search, +5 more.
task_reassignA
Release a pending OR running task back to unclaimed so another agent can take it over. Transitions running→pending (or keeps pending), clears claimant, and revokes the attempt token while sealing it into a handoff generation (stale_attempt_ids + reassigning=true, ≈dsh handoffId): any later complete/fail/approve by the old implementer — with its old token OR even tokenless during the reassignment window — is rejected by staleAttemptRejected. The new owner claims it to start a fresh attempt and clear the handoff state. Use when the current implementer is stuck/lost/gave up and you want an uncontested take-over. Terminal tasks refuse (can't resurrect).
Inputs: task_id (required), note.
task_resumeA
恢复一个 interrupted/failed/superseded 任务:复用原任务的描述(prompt)+session_id(若有)+workdir+agent,重新派发给同一 worker 续跑。session_id 复用实现真 resume(claude/codex/qwen 保留上下文);DSH/opencode 无 resume 则全新跑。复用原 task_id,保留 workflow/dependencies 挂链。可用 args.prompt/session_id/workdir/model 覆盖。终态 completed 不可恢复。
Inputs: task_id (required), agent, model, prompt, +5 more.
task_sedimentB
任务完成自动沉淀:把一条已完成/已有结果的任务(标题+描述+结果)提炼成知识点写进向量记忆(memory_add)。供后续任务 memory_search 复用。
Inputs: task_id (required), cwd, scope, source, +1 more.
task_supersedeA
Mark a task superseded by another (decision changed: implementer swapped or a step cut). Transitions any non-terminal status→superseded, records superseded_by (task id) + reason. Use this instead of completed/failed when a task didn't finish but was overtaken — keeps the task log honest (no zombie pending tasks).
Inputs: reason (required), task_id (required), superseded_by.
vision_analyzeA
Bare-API vision analysis of an image via vision-LLM (Qwen3-VL-235B-A22B-Instruct on <LLM_ENDPOINT_QWEN>). For image recognition / scene analysis / reading legible text identifiers. Pass either image_path (local file) or image_data_url (data:...;base64) — not both. Returns a Chinese analysis report. NOTE: this is a scene/text-visibility analyzer, NOT a barcode decoder — high-density barcode VALUES must be decoded by a real decoder (ZXing/Dynamsoft), the VLM cannot. Key read at runtime from opencode.json qwen provider.
Inputs: model, prompt, image_path, timeout_sec, +1 more.
worker_globC
Glob files within worker's authorized workdir parent tree.
Inputs: pattern (required), workdir (required).
worker_grepC
Search file contents (ripgrep) within worker's authorized workdir parent tree. Returns file:line:match.
Inputs: pattern (required), workdir (required), max, glob.
worker_list_dirC
List directory entries (one level) within worker's authorized workdir parent tree.
Inputs: workdir (required), path.
worker_read_fileA
Read a file the worker is authorized to access (bounded by its workdir, NOT BRIDGE_WORK_ROOT). For spawned worker sessions only. Worker passes file_path relative to its workdir, or absolute within workdir parent tree.
Inputs: workdir (required), file_path (required), limit, offset.
workflow_evolveA
【自适应重规划引擎】总入口:封装六个 DAG 演进动作(fork 备选 agent / append 后置追加 / insert 前置插入 / degrade 降级验收 / rollback 阶段回退 / branch 条件分支),并【服务端强制】演进护栏——演进计数 ≤3(超限拒)、演进 ≥2 次强制独立审闸(未过审 gate_required 拒执行)、不破坏已完成段(rollback 是显式豁免:授权推翻已完成段)、留痕统一记进 workflow 元数据。调用方只要给 action+目标任务+理由,护栏/计数/留痕/面板标记引擎代管,不必手动拼 task_fork/task_depend/task_create。
Inputs: action (required), reason (required), workflow_id (required), score, +10 more.
workflow_planA
Orchestrator 自动拆解(L1):给一句话目标, 自动产结构化的多段 DAG 计划。调用拆解器模型把 goal 拆成 stages[](每段 title/description/criteria/agent/depends_on),做结构自检(数组非空/每段可验收 criteria/agent 已注册/依赖引用合法/无环反向DFS/≤8段),并按复杂度定 L3 人审闸:复杂(段数或跨 agent 并行超阈值)→approval_required=true 交主控审核;小任务→auto-approve 可直接接 workflow_start。不直接建任务, 返回可执行计划供人审/改造。拆解器经 runAgent(继承 429 退避/多模型轮转)。
Inputs: goal, title, stages, context, +4 more.
workflow_startA
把一次多Agent工作流落成阶段依赖链任务 DAG,供面板「当前工作流」按卡片/聚焦图展示。支持四种范式:①线性链模板 'bmad'(需求→架构→实现→评审) 或 args.stages 自定义任意阶段;②竞争式范式 args.paradigm='compete'(同一问题派 ≥2 个视角并行各出方案 → 主控收敛最优,传 args.competitors:[{agent,view}],可 args.converge_title/args.converge_agent);③合作式范式 args.paradigm='collaborate'(设计→实施→审核 三段链且实施者与审核者分离,传 args.designer/args.implementer/args.reviewer,默认 claude→codex→claude);④动态路由范式 args.paradigm='dynamic'(自动分析 goal 选 compete 或 collaborate,路由理由写进工作流 meta + 各 task description「上墙」,参数缺省时用默认 agent 分配:compete=claude/codex/qwen 三视角,collaborate=claude设计/codex实施/qwen审核)。 模板库:template 可选 bmad/bmad-lite/review-only/fix-flow,或 '_list' 返回清单; 并行阶段:args.stages[i].agents=[a,b,c] 或 parallel:true → 该阶段铺 N 个同深度并行任务(共享前置,下一阶段依赖该阶段全部)。args.workflow_id 稳定 id + args.title 让多次调用归并成同一张工作流卡;可选 args.approve_each_phase 每阶段人工门。返回阶段 task_id 链。
Inputs: goal, title, stages, designer, +8 more.
Tool Schema Changelog
Recent tool additions, removals, and schema changes observed during successful MCP inspections.
Loading tool schema changelog…
TDQS
Scored across 63 tools
Several tools overlap in purpose: run_codex/run_claude/run_qwen/etc. duplicate agent_invoke with a name parameter; read_file/list_dir mirror worker_read_file/worker_list_dir; multiple messaging read tools (inbox_read, inbox_wait, bus_history) share retrieval semantics. Boundaries are often defined only through context (worker vs bridge), so an agent can easily misselect.
All names use snake_case, but the ordering convention is inconsistent: many tools follow domain_action (task_create, memory_search), while others use verb_noun (read_file, list_dir, run_codex). Prefixes are mostly grouped by domain, but mixed verb/noun order reduces predictability.
63 tools are far beyond the recommended 3-15 range and exceed the 50+ extreme-mismatch threshold; many tools (e.g., six run_* wrappers, duplicate file APIs, parallel memory and messaging interfaces) could be consolidated. The server’s complexity may justify breadth, but the count is excessive and burdens selection.
The surface covers task lifecycle, agent management, workflows, messaging, vector memory, file ops, checkpoints, and stats, leaving few functional gaps. Minor missing operations like task update/edit or bridge-side file write are workable via agents or task forking.
Maintenance
Related MCP Connectors
Task manager for AI agents. Orchestrate agents to achieve large goals in unified control plane.
- aXDeprecatedio.github.ax-platform
Agent-native collaboration network: orchestrate a team of long-running agents from any MCP client.
AI work orchestration for plans, tasks, teams, and coding-agent dispatch.
Design, build and run teams of AI agents with checked handoffs between them.
Related MCP Servers
- FlicenseNot gradedqualityNot gradedmaintenanceA production-grade coordination hub that enables AI agents and human teams to work as a single organism by sharing tasks, context, decisions, and persistent memory across projects. It features two-tier agentic memory with per-agent hot caches, inter-agent messaging, and multi-agent authorship tracking for seamless collaboration.2-
- AlicenseNot gradedqualityDmaintenanceEnables multiple LLM agents across devices to form teams, share knowledge, memory, and tasks with live status via a web dashboard and distributed-systems reliability.Apache 2.0
- AlicenseNot gradedqualityBmaintenanceConnects multiple AI agents (Codex, Claude Code, Grok, etc.) into a coordinated, auditable team with shared memory, task ownership, and human oversight, enabling parallel collaboration without conflicts.239Apache 2.0
- AlicenseNot gradedqualityCmaintenanceEnables multiple AI coding CLIs (Claude Code, Gemini/Antigravity, Codex, and OpenCode) to collaborate as a coordinated team by routing cross-agent prompts, sharing messages and review tickets, tracking tasks on a shared store, and isolating each agent in its own Git worktree with turn-budget safeguards—all inspectable and steerable from a local web dashboard.MIT