Codex Team Workbench
Local MCP server that maintains a shared task board and team map for Codex collaborators — tracking tasks, members, workflows, runtime observations, and source-labeled history — without spawning or controlling agents.
Open the persistent team panel/board UI, optionally selecting a board.
Create, archive, or restore boards; bind them to real native session UUIDs with confirmed automatic visible-text sync.
Create and update shared tasks with dependencies: claim, complete, reopen, edit, reassign, soft delete, all with revision checks.
Record members (name, status, parent, native/runtime IDs) and workflow groups for team-map grouping.
Record runtime states observed via collaboration
list_agents/wait_agent, rejecting stale observations.Request native execution and host actions (navigate/resume), then record their real, source-labeled results.
Append and page recorded work observations for a member or workflow.
Sync or import bounded, paginated visible thread history — never reasoning, tool bodies, or credentials.
Resolve child agent UUIDs from relayed
subAgentActivityitems.Read board snapshots, including task revisions and source-labeled observations.
Enable or disable per-board background history sync after explicit user confirmation.
Read-only in several tools; cannot create/resume agents, read private host sessions, or perform host operations itself.
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., "@Codex Team Workbench Create a task for the login API refactor and assign it to Alice"
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.
Team Workbench / 团队工作台
A local MCP workbench for native Codex collaborators, inspired by DeepSeek Harness. Tasks, lifecycle observations, exact parent relationships and source-labeled records share one conversation panel.
0.5.6 新增交互团队画布、真实任务完成率和活动卡片,并改进插件更新后的界面识别。
画布支持缩放、适应、100% 复位与拖动平移;刷新保留视图。
圆形、方形与菱形区分普通成员、工作流分组和 Hook 来源观察。
任务概览由真实已记录任务计算完成率;活动动效只在新记录到达时出现。
发布使用独立的 0.5.6 UI 资源地址和 MCP 页面修订标记。
0.5.5 的同步初始化、稳定读取目录与监视状态恢复修复继续保留。
Linux 原生 Hook 与后台读取链路已在一个授权安装范围内验证;标准安装仍需宿主正常
审阅 Hook。GitHub 源码发布、客户端安装和实际连接重载分别处理。原稳定源码可追溯至
3a66a9a。
已实现与实测
新建时可一并关联准确的主会话并启用已确认的自动同步;身份暂缺时显示“待关联”, 主代理登记真实身份后自动接上。可取消等待;子代理登记不会改变工作台根会话。
SQLite 持久化任务、负责人、DAG 依赖、原子领取与版本冲突检查。
原生
SubagentStart、SubagentStop、PreModelRequest回调写入插件自己的元数据 journal;普通 MCP 快照自动读取,无需反复发同步指令,不使用模型轮询状态。真实同一子代理的「运行→空闲→续接运行→空闲」已验证;续接运行由
PreModelRequest捕获。回合停止与任务成功分开记录。真实父子孙关系经公开线程父节点元数据核对;树形节点与工作流分组分别保留含义。 只有同一会话关联时,不虚构父子关系。
原生计时进程实际运行 135.169 秒。运行证据超过两分钟时保留「运行后未收到停止、 活跃性待确认」,真实停止回调到达后才转为空闲。
初始指令、执行结果与可见会话快照分开保存。快照是有来源、有边界的可见历史, 不冒充完整工具或推理转录。
已绑定会话的可见历史可通过公开分页接口直接读取;刷新运行记录直接读取已接收事件, 均不向主对话发消息。打开按钮使用宿主声明的
openLinks/ui/open-link,准确保留 拒绝或链接回执;当前客户端是否接受codex://仍待实测。明确对当前工作台启用后,后台监听现有插件事件日志,在子代理回合停止后自动同步 一页可见消息,按事件键去重,无需主代理回读。旧工作台默认关闭,不扩大 Hooks 范围。
启动、续接及可选的主会话协助操作明确标注会发送请求;任务正文保存在后端,执行与 续接消息只带请求标识。消息发出不等于执行成功;旧代理不在目录时不新建替身。
0.5.5 的初始化、状态恢复与原生读取回归见 验证说明。 新建和取消同步的具体行为见 初始化流程, 按钮路径和较早的原生验收见 0.5.2 按钮验收。 公开仓库仅含合成测试数据,不含真实会话标识、转录、任务数据库或测试回执。 后台架构、实测耗时与尚未测量的部分见 数据流说明。 0.5.2 隔离后端已用真实新代理及同一代理续接验证,两个停止事件到会话页保存分别为 1158 / 1021 毫秒。该验证使用现有测试探针;打包 Hooks 与客户端画面仍待实际加载验收。
Related MCP server: basic-mcp
从 GitHub 添加
标准目录清单位于 .agents/plugins/marketplace.json,./ 指向仓库根部插件。
市场名保持 codex-team-workbench,插件名保持 team-workbench。已经添加这个
GitHub 市场的用户应更新原来源,不需要重复创建插件。跟踪 main 可获取后续源码,
固定 commit 不会跟随更新;市场同步、安装更新与活跃 MCP 连接重载是不同步骤。
详见 市场与更新说明。
官方已核实的工作区同步路径需要工作区管理员:
打开 Admin → Plugins → Add → Import marketplace。
Source 填
https://github.com/912925635/codex-team-workbench;Path 留空; Branch 填main,或固定到一个审核过的完整 commit。导入后按工作区策略安装 Team Workbench,并正常审阅三项元数据 Hooks。
这是工作区 GitHub 来源导入,不是单纯连接 GitHub 应用。个人账号若没有该管理入口, 这里不假设存在同名按钮;使用客户端实际提供的原生安装入口。 官方 GitHub 导入说明 说明了管理员要求、仓库根路径和同步方式。
已有私有插件与新 GitHub 来源可能是不同身份,数据目录也可能不同。不要先卸载旧版, 不要删除旧数据库再重装;先确认身份、备份原数据并核对新实例的数据根。 仓库不嵌入任何人的私有插件 ID,也不自动把个人插件迁移到工作区。 详见 安装、保留数据与移除。
Hooks 与权限
源码声明三项待审阅的插件 Hooks。发布或克隆仓库不会修改本机 hooks.json、信任记录、
模型设置或现有安装。安装后仍需在 Settings → Hooks 或输入框的 Review hooks
中正常审阅;未受信任时不执行。
处理器只持久化明确绑定、未归档工作台对应根会话的事件/代理/回合标识、类型和时间。 不保存提示词、对话正文、工具载荷、凭据或 Token 数据;不读宿主私有会话目录或数据库。 启用这个范围需要用户正常批准,单个测试会话的授权不等于所有工作台的授权。 每次回调使用短时 Node 进程,没有常驻监听器或远程数据上传。
若已有开发探针,先确认插件自身回调正常,再只移除准确匹配的开发定义。 工作台不要同时读取两份 journal。迁移脚本只生成计划,不自动改配置。
本地开发
需要已有 Node.js 24+(使用内置 node:sqlite);Windows 公开元数据读取还需要
已有 PowerShell 7。没有 npm 依赖,不需要 API Key 或另装 Codex CLI。
0.5.4 的 Linux 适配改用 Node 读取已有原生 Unix socket,不依赖 PowerShell;
服务器条件、诊断和实测边界见 Linux 同步说明。
npm run check
npm test
npm run preview预览仅绑定 127.0.0.1 动态端口,使用真实本地存储;它不是桌面插件安装或 UI 验收。
Ctrl+C 只停止该预览。node src/server.mjs 提供 stdio MCP。
可用 npm run pack:local 生成源码包;检查、测试和打包脚本不会安装插件、信任 Hooks、
重启宿主或启动模型。
数据根优先级:TEAM_WORKBENCH_DATA,否则 PLUGIN_DATA/team-workbench,否则当前
工作目录下的 .team-workbench。升级必须核实实际根目录;相同包名不保证跨来源数据迁移。
当前边界
新 GitHub 安装的面板激活、视觉连线、点击跳转和续接按钮的完整客户端验收仍未完成。 公开导航工具实际返回过成功,但这不等于目视确认界面。当前会话的 MCP 热重载也未核实, 关闭重开面板不保证替换旧连接。
插件不能凭匿名 metadata 自动取得原生线程 UUID;绑定必须来自明确授权的准确来源。 代理目录丢失后的历史代理恢复尚未证明可行。续接依赖实际父会话能够访问那个代理及 受支持的协作工具;不向 V2 子代理直接发送 app-server 输入,不偷偷换成自管团队。
许可证
MIT。DeepSeek Harness 的任务板设计与 部分语义适配保留 MIT 许可和归属,见 THIRD_PARTY_NOTICES 及 上游许可证。
Available Tools
21 toolsteam_append_recordB
Append an explicitly agent-reported, user-visible work record, not a native transcript. Use a unique recordId for idempotent retry. Do not collect hidden reasoning, credentials, or private session files. All content stays labeled reported.
Inputs: role (required), text (required), nodeId (required), boardId (required), +4 more.
team_archive_boardA
Archive or restore a workbench after explicit user confirmation of scope. This is reversible; it never deletes tasks, member records, transcripts or project files and does not stop agents.
Inputs: action (required), boardId (required), confirmed (required), expectedRevision (required).
team_bind_agent_activityB
Resolve and register one recorded child using exact subAgentActivity path/UUID items relayed from its recorded spawning parent. Does not read host history itself. Never substitutes another parent or guesses a UUID. Returns nextRead only after unique mapping; missing/ambiguous/unavailable sources are stored as visible resolution states.
Inputs: boardId (required), memberId (required), expectedRuntimeRef (required), activities, +4 more.
team_bind_sessionB
Bind a full native UUID; automaticHistory=true with confirmed=true also initializes this board scoped visible-text sync in the same call. Changing an existing root still requires exact previousSessionId, confirmRebind=true and user-reviewed old/new IDs and history isolation. Old sources are preserved. Never guess an ID.
Inputs: boardId (required), sessionId (required), expectedRevision (required), evidence, +5 more.
team_create_boardC
Create a local workbench. Supply the actual originating sessionId to bind it immediately. For user-approved automatic sync, set automaticHistory=true and confirmed=true; if identity is not yet known, registering the real lead UUID completes initialization automatically. Anonymous session metadata is never a thread UUID. Does not create any agent or thread.
Inputs: title (required), evidence, confirmed, sessionId, +2 more.
team_create_taskA
Create an unassigned pending task with optional dependencies. Task assignment alone does not dispatch work to any agent.
Inputs: title (required), boardId (required), blockedBy, description.
team_get_requestARead-only
Read one exact execution/host-action request with its stored instructions and required target/parent revisions. Returns no unrelated tasks, messages or work history. Use for compact panel requests instead of loading an entire board.
Inputs: kind (required), boardId (required), requestId (required).
team_import_thread_pageB
Relay a bounded page actually returned by the public codex-app-tools.read_thread host tool for an explicitly bound member. Starts no model and calls no host API itself. Exact threadId binding required. Import selected visible messages/status/tool summaries, never reasoning or credentials. Snapshots remain marked agent-relayed, paginated/possibly truncated, not complete native transcripts.
Inputs: turns (required), readAt (required), boardId (required), evidence (required), +6 more.
team_observe_runtimeB
Record an actual collaboration list_agents or wait_agent state for the exact recorded runtime path. Evidence and observation time are lead-relayed and not server-authenticated; stale observations cannot overwrite newer ones. resumeAvailable is only true when the lead actually sees the exact agent and has the supported followup tool.
Inputs: state (required), source (required), boardId (required), evidence (required), +6 more.
team_panel团队工作台B
Open the persistent shared task board and team map. With boardId, remember that selection if the host supplies anonymous session metadata; {} reopens it. This correlation is not a native Codex thread ID. Does not discover or control agents.
Inputs: boardId.
team_read_recordsBRead-only
Page recorded work observations, relayed public-host snapshots and execution receipts for one node. Does not directly read native threads or private host sessions/databases. Complete-native access is explicitly distinguished from stored bounded snapshots.
Inputs: nodeId (required), boardId (required), nodeKind (required), after, +1 more.
team_record_executionB
Record a real lead spawn receipt, worker start/result, or parent-collected result. These are reported observations, not authenticated runtime telemetry. Never invent runtimeRef or success. confirm completes the task only after worker report and collection.
Inputs: runId (required), action (required), boardId (required), evidence (required), +9 more.
team_record_host_actionB
Record one exact supported host result with its actual source. ui/open-link acknowledgment is accepted, not visual success; ui/message delivery alone cannot succeed. User-confirmed cancelled only closes the local workbench wait and never sends a message or stops a native operation.
Inputs: result (required), source (required), status (required), boardId (required), +4 more.
team_record_workflowB
Record an actual caller-owned workflow/run group, with source evidence and revision. This groups agents visually and does not replace their actual parent relationships or start any work. Use expectedRevision=0 for a new group.
Inputs: title (required), status (required), boardId (required), evidence (required), +3 more.
team_report_memberB
Record a caller-reported member. An explicitly sourced native UUID for lead also establishes the first unbound board association; user-approved pending automatic sync then initializes. Child reports never bind the board and existing roots never change. Does not create, resume, or control a native agent.
Inputs: name (required), status (required), boardId (required), memberId (required), +10 more.
team_request_executionB
Request native execution through the lead. Starts no agent; the lead must use actual host collaboration tools and record their returned reference.
Inputs: taskId (required), boardId (required), expectedTaskRevision (required), workflowId, +2 more.
team_request_host_actionB
Record a user-requested exact-target navigation or supported continuation. Does not execute the host operation. A message acknowledgment is not success.
Inputs: kind (required), boardId (required), memberId (required), instructions, +1 more.
team_set_history_syncA
After explicit user confirmation, enable/disable automatic bounded visible-text sync for this exact board root. An empty expectedRootSessionId is allowed only to cancel an unbound board pending initialization. Existing history is retained; no native agent or Hook configuration is changed.
Inputs: boardId (required), enabled (required), confirmed (required), expectedRevision (required), +1 more.
team_snapshotCRead-only
List boards or read one board, including task revisions and source-labeled observations. Ingests configured plugin-owned Hook metadata and schedules only confirmed per-board background history. No model invocation or native agent control.
Inputs: boardId.
team_sync_native_historyA
Read and save one bounded visible-message page directly via public app-server thread/turns/list. Exact bound root, member UUID and fresh parent metadata required; never resumes a thread or sends a main-chat message. Older-page cursors must refer to an existing direct page of this same binding. No private host files, reasoning or tool bodies.
Inputs: boardId (required), memberId (required), expectedNativeId (required), beforeRecordId.
team_update_taskADestructive
Update a shared task with expectedRevision. Refresh after STALE_REVISION. actorId declares a workflow role, not host identity. Supports claim, release, complete, reopen, edit, dependencies, reassign, soft delete.
Inputs: action (required), taskId (required), actorId (required), boardId (required), +5 more.
Tool Schema Changelog
Recent tool additions, removals, and schema changes observed during successful MCP inspections.
Loading tool schema changelog…
TDQS
Scored across 21 tools
Several tools overlap heavily in intent: record_execution, record_host_action, record_workflow, and append_record all record related observations, while request_execution and request_host_action both create requests. Sync/import/read snapshot tools are also easy to confuse despite lengthy descriptions.
All tools use a consistent team_ prefix and snake_case, mostly following verb_noun. Minor deviations like team_snapshot and team_panel are noun-only rather than action-oriented.
21 tools is on the heavy side for a single workbench surface. The complexity of the domain partially justifies it, but the count contributes to selection difficulty.
The surface covers board lifecycle, task creation/update, member/execution records, history sync, and thread import. Some gaps remain around direct task listing/getting and explicit member management, but core workflows are covered.
Maintenance
Related MCP Connectors
Shared task layer for AI agents: tasks, handoffs and decisions that persist across tools.
Durable background job execution, async task scheduling, and state persistence for AI agents.
Persistent memory, knowledge graph, task management and more for AI agents via MCP.
Private-by-default, local-first memory/context/task orchestrator for MCP apps and agents.
Related MCP Servers
- AlicenseNot gradedqualityCmaintenanceEnables coding agents to persist and resume shared working state—tasks, progress, decisions, commands, tests, checkpoints, and handoffs—backed by SQLite while keeping Git authoritative for files and repository state.MIT
- AlicenseNot gradedqualityCmaintenanceEnables task management with persistent SQLite storage, along with tools, resources, subscriptions, prompts, and completions over stdio or Streamable HTTP with per-user authentication.MIT
- AlicenseNot gradedqualityBmaintenanceEnables AI agents like Codex, Grok Build, and Antigravity to share persistent local memory, preferences, project context, tasks, and handoffs using Markdown and SQLite.1MIT
- FlicenseNot gradedqualityAmaintenanceEnables AI coding agents to read and write persistent local memory during sessions, storing and semantically searching project decisions, preferences, error patterns, guidelines, and cross-project standards from SQLite stores on the user's machine. An optional companion server also exposes an autonomous task queue for executing queued work.-