ClickUp MCP Server
Enables multi-user access to ClickUp workspaces, tasks, comments, custom fields, time tracking, and more through 150 tools, with per-user OAuth authorization.
README
ClickUp MCP Server
A multi-user MCP server for ClickUp. Each person who connects authorizes with their own ClickUp account, and every API call runs against their OAuth token — so ClickUp enforces their real permissions and no one can see anything through this server that they could not see in the ClickUp UI.
Covers 150 tools across ClickUp API v2 and v3: the Workspace hierarchy, tasks, comments, custom fields, time tracking, tags, checklists, dependencies, attachments, Docs, Chat, views, goals, templates, webhooks, and Workspace admin.
Why this one is different from the other servers in this estate
The OAuth servers here (Whoop, Withings, Fitbit) hold one token set in a module-level singleton — whoever authenticates last owns the server. That is fine for personal health data and wrong for a team on ClickUp.
This server keeps a per-user grant. The MCP access token presented on each request is mapped to exactly one ClickUp authorization:
Claude → /authorize → app.clickup.com consent → /clickup-callback
→ ClickUp token + GET /v2/user (identity) → grant row (encrypted)
→ MCP code → MCP access + refresh token, both bound to that grant
→ tool call → get_access_token() → grant → ClickUp API
The response cache, the rate-limit bucket, and the audit trail are all namespaced by grant id.
Quick start
-
Create a ClickUp OAuth app at https://app.clickup.com/settings/apps. Set the redirect URL to exactly
{SERVER_URL}/clickup-callback. -
Configure
cp .env.example .env # generate a key for the token store python -c "from cryptography.fernet import Fernet; print(Fernet.generate_key().decode())"Fill in
CLICKUP_CLIENT_ID,CLICKUP_CLIENT_SECRET,SERVER_URL,TOKEN_ENCRYPTION_KEY, andCLOUDFLARE_TUNNEL_TOKEN. -
Run — either build from source:
docker compose up -d --build curl https://your-host/health...or run the published image, which is what
docker-compose.ymlpoints at by default:docker compose pull && docker compose up -dPublished to
ghcr.io/npab19/clickup-mcpforlinux/amd64andlinux/arm64on every push tomain, and pullable without authentication. Pin a specific build withCLICKUP_IMAGE_TAG=sha-<commit>in.env.Because the compose file declares both
image:andbuild:,upwill use a locally-built image if one is tagged — rundocker compose pullto fetch the published one, ordocker compose up -d --buildto force a local build. -
Connect your MCP client to
https://your-host/mcp. It will walk you through the ClickUp consent screen on first use. Each additional user repeats step 4 and gets their own grant.
Configuration
| Variable | Default | Purpose |
|---|---|---|
CLICKUP_CLIENT_ID |
— | Required. OAuth app client id. |
CLICKUP_CLIENT_SECRET |
— | Required. OAuth app client secret. |
SERVER_URL |
http://localhost:8000 |
Required in production. Must be HTTPS; the process refuses to start on a non-HTTPS non-localhost URL. |
TOKEN_ENCRYPTION_KEY |
— | Required. Fernet key encrypting ClickUp tokens at rest. |
CLICKUP_DB_PATH |
/data/clickup.db |
SQLite store. |
CLICKUP_TOOL_PROFILE |
core |
core = phases 1–2 (89 tools). full = all 150. |
CLICKUP_ENABLE_DESTRUCTIVE |
false |
Expose the 20 tools that irreversibly delete content. |
CLICKUP_ADMIN_EMAILS |
(empty) | Comma-separated ClickUp emails allowed the 20 admin tools. Empty means nobody. |
CLICKUP_RATE_CAPACITY |
60 |
Per-user token bucket size. |
CLICKUP_RATE_REFILL_PER_MINUTE |
60 |
Per-user refill rate. |
On the tool profile
150 flat tools measurably degrades a model's tool selection — past roughly 60–80,
mis-picks rise and the tool list alone eats context. Everything is built; the
profile controls what is advertised. core is the recommended default and covers
the daily surface: hierarchy, tasks, comments, custom fields, time tracking, tags,
checklists, dependencies, attachments.
Tools by phase
| Phase | Content | Tools |
|---|---|---|
| 1 — core hierarchy + tasks | workspaces, spaces, folders, lists, tasks, comments, custom fields, template listing, move-task | 49 |
| 2 — time tracking + collaboration | time entries, tags, checklists, dependencies, members, attachments, view reading | 40 |
| 3 — API v3 | Docs, Chat, entity attachments, ACLs | 31 |
| 4 — admin | view authoring, goals, webhooks, users, guests, groups, audit logs | 30 |
Tools are filed by what they are for, not by which API version serves them —
move_task_to_list is v3-only but is core task work, so it sits in phase 1.
search_tasks (ClickUp's GetFilteredTeamTasks) is the workhorse — one call
against a whole Workspace with filters, instead of walking
Spaces → Folders → Lists → Tasks.
Safety model
Four independent layers, each fail-closed:
- ClickUp's own permissions. The real boundary. Every call carries the caller's token, so the server cannot exceed what that person can do in the UI.
- Tool gating (
policy.py). Phase, destructive, and admin gates decide what each caller is shown. A hidden tool is neither advertised nor callable. - Confirmation guards. Every destructive tool takes
confirm: bool = Falseand refuses without it. ClickUp deletes cascade — a Space takes its Folders, Lists, and Tasks with it. - Audit (
audit.py). Everytools/callis recorded with the acting ClickUp user, arguments, outcome, and duration.
Secrets: MCP tokens are stored as SHA-256 hashes; ClickUp tokens are Fernet-encrypted
because they must be replayed upstream. ClickUp access tokens never expire, so a
leaked store is permanently valuable — treat the clickup-data volume as a secret.
ClickUp API notes
Things that differ from a typical OAuth integration, all verified against the live API:
- No PKCE, no scopes. ClickUp's authorize endpoint is
https://app.clickup.com/apiand supports neither. The user picks which Workspaces to grant on the consent screen. PKCE between your MCP client and this server is unaffected. - Tokens never expire and no refresh token is issued. There is no refresh path
in
client.pyby design. A 401 means the user revoked the integration; the only recovery is re-authorization, so the server says so instead of retrying. - Rate limits are per token, i.e. per user: 100 req/min on Free, Unlimited, and Business; 1,000 on Business Plus; 10,000 on Enterprise.
- The v3 spec is only a fragment — Chat, Docs, and audit logs. Tasks, Lists, Spaces, time tracking, and everything else are v2 only. Both specs are vendored in this repo.
Development
pip install -e ".[dev]"
pytest -q # 367 tests
tests/test_isolation.py is the gate: it proves the response cache cannot serve one
user's data to another, that each request carries the right token, and that an MCP
token refresh preserves its ClickUp binding. Run it before touching store.py,
client.py, or oauth_provider.py.
Regenerating requirements.lock — do it in a Linux container, not on Windows.
pip-compile resolves for the host platform and will otherwise pin pywin32,
which breaks the Docker build:
docker run --rm -v "${PWD}:/w" -w /w python:3.12.10-slim `
sh -c "pip install -q pip-tools && pip-compile --quiet --strip-extras --output-file requirements.lock pyproject.toml"
Layout
src/clickup_mcp/
runtime.py # singletons + startup config checks
store.py # SQLite: clients, codes, tokens, grants, audit
oauth_provider.py # OAuth 2.1 server delegating to ClickUp
client.py # per-grant HTTP client, cache, rate-limit handling
context.py # MCP token -> ClickUp grant (the multi-user hinge)
policy.py # per-identity tool gating
audit.py # tool-call audit trail
rate_limit.py # per-identity token bucket
transform.py # response slimming
validation.py # input coercion + confirm guard
app.py # FastMCP instance + @tool decorator
server.py # callback route, /health, entrypoint
tools/ # 20 modules, 150 tools
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。