clickup-custom-mcp
An MCP server for interacting with ClickUp via its REST API, providing tools to manage tasks, search work, and explore workspace structure.
README
clickup-custom-mcp
A stdio MCP server that talks to the ClickUp REST API v2 directly.
It exists because ClickUp's hosted MCP (mcp.clickup.com) caps a Free workspace at
50 calls per 24 hours and refuses personal API tokens, while the same account's REST
API allows 100 requests per minute. This server trades a daily ceiling for a
per-minute one, and keeps the token on your machine instead of in a hosted integration.
Requirements
- Node 20+ (see
enginesinpackage.json) - A ClickUp personal API token
Setup
npm install
npm run build
Generate a token in ClickUp: avatar → Settings → Apps → API Token → Generate.
Copy .env.example to .env at the repo root (.env is gitignored) and fill in the token:
cp .env.example .env
The server loads that file itself via process.loadEnvFile, so the token never has to be
written into Claude Code's config — where claude mcp add -e would store it in plaintext.
Anything already present in the environment wins over .env, and if the file is missing
(or the Node build has no loadEnvFile) the server simply falls back to plain environment
variables.
Environment variables
| Variable | Required | Purpose |
|---|---|---|
CLICKUP_API_TOKEN |
yes | Personal API token. The server exits with a message if it's unset. |
CLICKUP_TEAM_ID |
no | Pin a workspace. Without it the first workspace the token can see is used, and a note goes to stderr if there is more than one. |
CLICKUP_RATE_LIMIT |
no | Requests/minute the client allows itself. Defaults to 90 — deliberate headroom under ClickUp's 100. |
Register with Claude Code
claude mcp add clickup-custom-mcp -- node /absolute/path/to/clickup-custom-mcp/dist/index.js
No -e CLICKUP_API_TOKEN=... is needed — the .env file covers it.
Run it standalone with npm start. stdout is the JSON-RPC channel, so every diagnostic
(including a per-request log line with ClickUp's remaining quota) goes to stderr.
Tools
Six coarse tools, each shaped around a job rather than an endpoint. The fan-out lives in
the server, not in the model: one get_my_work call costs 2 HTTP requests where walking
space → folder → list → task from the model side costs thirty.
| Tool | What it does |
|---|---|
get_workspace_map |
Spaces, folders, lists (with their status names) and members, plus your own numeric user id. Call it first when you need an id for anything else. Cached 15 minutes; optionally includes custom field definitions. |
get_my_work |
Every open task assigned to you across the workspace, grouped by overdue / today / this week / later. The one for a standup or a backlog sweep. |
search_tasks |
Filtered task query across the whole workspace in one request — lists, spaces, assignees, statuses, tags, due/updated date ranges, ordering. Paginates automatically. |
get_task |
Full detail for one task: description, custom field values, subtasks, optionally comments. |
create_task |
Create one or many tasks in a single call. Each is reported individually; one failure does not abort the rest. |
update_task |
Update one or many tasks — status, assignees, due dates, priority, renames, archive. Also reported individually. |
Every tool reports how many HTTP requests it spent, and any result that hit the page cap says Truncated rather than quietly returning a short list.
Design notes
- Rate gate (
src/clickup.ts) — a sliding 60-second window plus a concurrency cap of 4. A fixed-interval refill would let a burst of 90 fire in the first second and then stall for 59; tracking actual send times spreads the same budget. - 429 handling — backoff is driven by ClickUp's own
X-RateLimit-Resetheader, retried up to 3 times. The client also coasts down whenX-RateLimit-Remainingdrops to 2 rather than taking a 429 mid-fan-out. - Structure cache — hierarchy, custom fields and member lists are memoised for 15 minutes. Task data is deliberately never cached.
- Compact output (
src/format.ts) — raw ClickUp task JSON runs 2–4KB each. Every tool projects down to the fields a person actually reads and renders markdown instead of JSON.
Layout
src/index.ts entrypoint: env loading, client + server wiring, stdio transport
src/clickup.ts REST client: rate gate, retry/backoff, TTL cache, task pagination
src/tools.ts the six MCP tool definitions and their fan-out
src/format.ts task projection and markdown rendering
推荐服务器
Baidu Map
百度地图核心API现已全面兼容MCP协议,是国内首家兼容MCP协议的地图服务商。
Playwright MCP Server
一个模型上下文协议服务器,它使大型语言模型能够通过结构化的可访问性快照与网页进行交互,而无需视觉模型或屏幕截图。
Magic Component Platform (MCP)
一个由人工智能驱动的工具,可以从自然语言描述生成现代化的用户界面组件,并与流行的集成开发环境(IDE)集成,从而简化用户界面开发流程。
Audiense Insights MCP Server
通过模型上下文协议启用与 Audiense Insights 账户的交互,从而促进营销洞察和受众数据的提取和分析,包括人口统计信息、行为和影响者互动。
VeyraX
一个单一的 MCP 工具,连接你所有喜爱的工具:Gmail、日历以及其他 40 多个工具。
graphlit-mcp-server
模型上下文协议 (MCP) 服务器实现了 MCP 客户端与 Graphlit 服务之间的集成。 除了网络爬取之外,还可以将任何内容(从 Slack 到 Gmail 再到播客订阅源)导入到 Graphlit 项目中,然后从 MCP 客户端检索相关内容。
Kagi MCP Server
一个 MCP 服务器,集成了 Kagi 搜索功能和 Claude AI,使 Claude 能够在回答需要最新信息的问题时执行实时网络搜索。
e2b-mcp-server
使用 MCP 通过 e2b 运行代码。
Neon MCP Server
用于与 Neon 管理 API 和数据库交互的 MCP 服务器
Exa MCP Server
模型上下文协议(MCP)服务器允许像 Claude 这样的 AI 助手使用 Exa AI 搜索 API 进行网络搜索。这种设置允许 AI 模型以安全和受控的方式获取实时的网络信息。