wb-mcp exposes wb-cli as MCP tools for AI workbench read/write operations.
Read/query across 49 tables via wb_query (search, growth, canvas, logs, decisions, people, doctor).
Manage todos/ideas/schedules/bugs/bookmarks via wb_todo and fine-grained tools (add, done, edit, tag, delete).
Create/edit notes via wb_note, wb_note_add, wb_note_edit.
Manage finance via wb_fin (accounts, transactions, budgets).
Admin tasks via wb_manage (AI-routed add, done, key management).
Drive dev-task pipeline via wb_dev_task: create, stage (9 steps), item checklists.
Read Obsidian cloud vault via wb_vault (list/search/read).
Read memory archive via wb_archive (list/show/gen/detect).
Upload local images to public URLs via wb_image_upload (per README; confirm required).
All writes require confirm:true; use args array for long text safety.
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., "@wb-mcp Show my pending todos"
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.
AI 工作台 wb-cli 的 MCP server 壳（stdio transport）。把 wb-cli 命令族暴露为标准 MCP tool， 供 Codex / DeepSeek CLI / Claude Code / WorkBuddy 等 MCP 客户端直连。
版本：6.3.0 · 由主仓 `scripts/release-wb-mcp.mjs` 一键同步发布
# 1. clone 本仓库（或从主仓跑 release 脚本生成）
git clone <repo-url> wb-mcp && cd wb-mcp
# 2. 安装唯一依赖
npm install
# 3. 验证（应输出 serverInfo name=wb-mcp 握手成功）
npm run probeRelated MCP server: Codex Harness MCP
WorkBuddy / Claude Code / Codex（stdio）
{
"mcpServers": {
"wb": {
"command": "node",
"args": ["<本仓库绝对路径>/scripts/wb-mcp.mjs"],
"env": { "WB_PROFILE": "workbuddy" }
}
}
}DeepSeek CLI/TUI
{
"mcp": {
"wb": {
"command": "node",
"args": ["<本仓库绝对路径>/scripts/wb-mcp.mjs"],
"env": { "WB_PROFILE": "workbuddy" }
}
}
}档位 | 权限 | 说明 |
workbuddy | 全权（admin key） | 默认；写操作仍需 confirm:true |
openclaw | 读写 | 部分写权限 |
hermes | 只读 | 双保险：只注册 wb_query tool + readonly key |
key 文件位置：`~/.workbuddy/agents/.env`（WB_API_KEY=wbk_xxx）。 新机器首次使用前，请在工作台网页「设置 → API Key」自助签发并写入该文件。
域 tool | 域 | 说明 |
wb_query | 只读调阅 | search/table/fill/growth/canvas/artifact/log/decision/orpt/people |
wb_todo | 待办闪念 | todo/capsule/idea/bug/book/schedule |
wb_note | 笔记 | note 命令族 |
wb_fin | 财务 | fin 命令族 |
wb_manage | 管理 | add/done/key |
wb_dev_task | 流水线 | dev-task 命令族 |
wb_vault | 云档只读 | vault list/search/read |
wb_archive | 记忆档案只读 | archive list/show/gen/detect |
另有细粒度 write tool，常用包括 wb_add、wb_note_add、wb_schedule_add 与 wb_image_upload。wb_image_upload 可把本机 PNG/JPG/GIF/WebP/BMP/AVIF 上传为公网图片链接。
适用场景：Agent 本地已生成或已下载一张图片，需要拿到公网 URL 用于分享、插入 Markdown 或写入笔记。
调用参数：
参数 | 必填 | 说明 |
path | 是 | Agent 本机图片文件绝对路径，例如 /tmp/workbench-image.png |
confirm | 是 | 写操作安全门，必须传 true |
调用示例：
{
"path": "/tmp/workbench-image.png",
"confirm": true
}成功返回的核心字段：
字段 | 用法 |
data.url | 公网图片 URL，直接发给用户或插入 Markdown |
data.markdown | 已拼好的 Markdown 图片片段 |
data.bucket / data.key | 排查存储对象用 |
data.size / data.contentType / data.sha256 | 完整性与类型校验用 |
Agent 使用规则：
不要编造或自行拼接 URL，必须使用本次返回的 data.url。
插入 Markdown 时优先直接使用 data.markdown。
path 必须是 Agent 本机可读文件；远程图片应先下载到本机临时文件再上传。
支持 PNG、JPG、GIF、WebP、BMP、AVIF，单文件 100MB；SVG 不支持。
返回 URL 无需登录即可访问，不得用于涉密图片。
// 字符串形态（传统）
{ command: "todo list --open" }
// args 数组形态（长文安全：正文原样传，不按空白切分）
{ command: "note add", args: ["note", "add", "这里500+字正文…", "#工作笔记"], confirm: true }写操作（add/done/edit/del/stage/item 等）必须传 `confirm: true`，否则被安全门拒绝。
三层写防线：confirm 意图门 → 串行队列 → 网关 RLS + key scope
敏感串（wbk_ 前缀 key、密码）发布前自动扫描，命中即中止
全类型录入（AI 判类路由到 待办/闪念/bug/日程/成长等表）。对应 CLI：add。参数：text*(录入原文（AI 自动判类；前缀「修复/优化/闪念/日程」可引导判类）), type(强制指定类型别名（跳过 AI 判类，如 todo/bug/idea/schedule）)。写操作须 confirm:true。
Inputs: text (required), type, confirm.
archive 记忆档案只读域（P4 · dt_mkq335）：archive list/show/gen/detect 命令族。list=档案清单（vault 桶 archive/ 前缀）；show=档案 md 全文（示例：archive show 顾铭）；gen=透传 A 段生成脚本；detect=近 N 天闪念+决策 × 档案的矛盾/关联检测（无 LLM 时自动降级素材摘要）。「顾铭眼中的刘总」档案是所有 AI 的公共记忆资产，会话开场可 archive list 查看、archive show 读取。只读安全，readonly profile 可见。
Inputs: command (required), args, confirm.
开发任务流水线域：dev-task 命令族（list/show/create/stage/item 九步链回写）。写操作需 confirm；并发打勾由串行队列保护。（本域含写操作）
Inputs: command (required), args, confirm.
新建开发任务（S0 需求汇总自动生成）。对应 CLI：dev-task create。参数：title*(任务标题), bugs*(关联待办 id（逗号分隔，支持前缀）)。写操作须 confirm:true。
Inputs: bugs (required), title (required), confirm.
S4 条目 / S7 清单打勾。对应 CLI：dev-task item。参数：id*(开发任务 id 前缀), key*(步骤（dev/test）), index*(条目序号（0 起）), status(目标状态), note(备注)。写操作须 confirm:true。
Inputs: id (required), key (required), index (required), note, +2 more.
九步链步骤回写（每完成一步必须回写）。对应 CLI：dev-task stage。参数：id*(开发任务 id 前缀), key*(步骤), status(步骤状态（缺省 done）), file(产物文件路径（内容写入步骤）), note(备注), kind(产物类型)。写操作须 confirm:true。
Inputs: id (required), key (required), file, kind, +3 more.
财务域：fin 命令族（账户/流水/分类/预算/目标/订阅等）。含写操作需 confirm。（本域含写操作）
Inputs: command (required), args, confirm.
直录 AI 点子（不经 AI 判类）。对应 CLI：idea add。参数：raw*(点子原文（含空格换行均可）), title(标题（缺省取原文前 30 字）)。写操作须 confirm:true。
Inputs: raw (required), title, confirm.
管理域：add（全类型录入 AI 判类）/ done（打标）/ key（API Key 签发/清单/吊销）。写操作需 confirm；key 域建议直接终端操作。（本域含写操作）
Inputs: command (required), args, confirm.
笔记域：note 命令族（笔记列表/新增/编辑）。含写操作需 confirm。（本域含写操作）
Inputs: command (required), args, confirm.
写笔记（capsules 表，#标签 自动解析）。对应 CLI：note add。参数：content*(笔记正文，含空格换行原样入库，#标签 自动解析)。写操作须 confirm:true。
Inputs: content (required), confirm.
改笔记（换正文重解析标签 / 仅改标题）。对应 CLI：note edit。参数：id*(笔记 id 前缀), file(新正文的 md 文件路径), title(新标题)。写操作须 confirm:true。
Inputs: id (required), file, title, confirm.
全域只读调阅：wb-cli 的 search/table/fill/growth/canvas/artifact/log/decision/orpt/people/doctor 命令族，跨 49 张表查询 AI 工作台数据（待办/闪念/日程/人脉/财务/成长/画布/日志/开发）；doctor=环境自检（node/key/网关/表权限/版本）。只读安全。
Inputs: command (required), args, confirm.
新增日程（自然语言或 --at 显式时间）。对应 CLI：schedule add。参数：text*(日程描述（自然语言）), at(显式时间 YYYY-MM-DD HH:mm), loc(地点)。写操作须 confirm:true。
Inputs: text (required), at, loc, confirm.
待办与闪念域：todo/capsule/idea/bug/book/schedule 命令族（列表/详情/完成/编辑/标签）。含写操作需 confirm。（本域含写操作）
Inputs: command (required), args, confirm.
删除待办（不可恢复，慎用）。对应 CLI：todo del。参数：id*(id 前缀)。写操作须 confirm:true。
Inputs: id (required), confirm.
完成待办。对应 CLI：todo done。参数：id*(待办 id（前缀匹配）)。写操作须 confirm:true。
Inputs: id (required), confirm.
编辑待办字段（至少传一个要改的）。对应 CLI：todo edit。参数：id*(id 前缀), title(新标题), priority(优先级), due(截止日期 YYYY-MM-DD), category(分类), status(状态)。写操作须 confirm:true。
Inputs: id (required), due, title, status, +3 more.
待办标签增删（至少传 add 或 del）。对应 CLI：todo tag。参数：id*(id 前缀), add(要加的标签，逗号分隔), del(要删的标签，逗号分隔)。写操作须 confirm:true。
Inputs: id (required), add, del, confirm.
vault 云档只读域（P1 · dt_iy2eek）：vault list/search/read 命令族，检索与读取 Obsidian 主库上云 md（COS vault-store 桶 obsidian/ 前缀 + vault_files 索引，刘总 5 年历史思考）。示例：vault search 精力 → vault read 。只读安全，readonly profile 可见。
Inputs: command (required), args, confirm.
Recent tool additions, removals, and schema changes observed during successful MCP inspections.
Loading tool schema changelog…
Scored across 20 tools
Several tools overlap in scope: wb_todo, wb_todo_done/edit/tag/del, and wb_add/wb_manage all cover todo creation/editing, while wb_dev_task duplicates wb_dev_task_create/stage/item. The read-only domains (wb_query, wb_vault, wb_archive) are clear, but the write-side boundaries are fuzzy enough to cause misselection.
Most specific operations follow wb_<domain>_<action> (wb_todo_edit, wb_note_add), and the wb_ prefix is consistent. However, broad domain tools use bare nouns (wb_todo, wb_fin, wb_vault) and some tools use verb-first or abbreviated forms (wb_query, wb_add, wb_manage, wb_del), creating a mixed convention.
At 20 tools, the server is at the high end of acceptable and feels heavy because many tools duplicate functionality covered by broader domain tools. A leaner set could drop or merge the generic wb_add/wb_manage with the specific add/done tools, but the count is not extreme.
Core workflows are covered: todo CRUD, note add/edit, schedule creation, idea capture, dev-task pipeline stages, and read-only vault/archive access. Minor gaps exist (no explicit note delete or dedicated schedule update/delete tool, finance operations hidden behind wb_fin), but agents can work around them.
MCP server for your apps' tools and custom tools, plus hosted AI agents and approval-gated workflows
Your org's AI agents, tasks, runs, search, and brain files as MCP tools and resources.
The Remote MCP server acts as a standardized bridge between LLM applications (like Claude, ChatGPT, and Cursor) and external services, enabling AI agents to access external tools and resources. Its primary capability is providing a centralized search tool to discover other MCP servers and their respective tools. Unlike local implementations, it runs remotely with OAuth authentication and permission controls for security.
Nifty's MCP server — exposes tasks, projects, messages, and files as tools for AI agents.
