
Plane MCP Server
MCPOfficial MCP for Plane, the open-source project management tool: lets agents read and manage projects, work items, cycles, modules and releases, with PQL queries. Use the hosted server with OAuth, or run locally via uvx with an API key; works with self-hosted Plane.
Plane MCP Server
A Model Context Protocol server for Plane. Gives an AI agent tools to read and manage projects, work items, cycles, modules, releases, customers and more.
Built on FastMCP and the official
plane-sdk.
- 30 tools, one per Plane resource, covering 207 operations
- Local or remote — stdio, streamable HTTP, SSE
- OAuth or API key authentication
Quick start
Get an API key from Plane: Workspace Settings → API tokens.
Add this to your MCP client's configuration:
{
"mcpServers": {
"plane": {
"command": "uvx",
"args": ["plane-mcp-server", "stdio"],
"env": {
"PLANE_API_KEY": "<your-api-key>",
"PLANE_WORKSPACE_SLUG": "<your-workspace-slug>"
}
}
}
}
uvx needs no install step. Requires Python 3.10+.
For a self-hosted Plane, add "PLANE_BASE_URL": "https://plane.example.com".
Transports
stdio — local
Runs as a subprocess of your MCP client. Configuration as shown above; needs
PLANE_API_KEY and PLANE_WORKSPACE_SLUG.
PLANE_API_KEY=... PLANE_WORKSPACE_SLUG=... uvx plane-mcp-server stdio
HTTP with OAuth — hosted
https://mcp.plane.so/http/mcp
The OAuth flow is handled on connect; no credentials in your config. For clients
without native remote MCP support, bridge with mcp-remote:
{
"mcpServers": {
"plane": {
"command": "npx",
"args": ["mcp-remote@latest", "https://mcp.plane.so/http/mcp"]
}
}
}
Requires Node.js 22+.
HTTP with a personal access token — hosted
https://mcp.plane.so/http/api-key/mcp
| Header | Value |
|---|---|
Authorization | Bearer <PAT> |
X-Workspace-slug | <workspace-slug> |
{
"mcpServers": {
"plane": {
"command": "npx",
"args": ["mcp-remote@latest", "https://mcp.plane.so/http/api-key/mcp"],
"headers": {
"Authorization": "Bearer <PAT>",
"X-Workspace-slug": "<workspace-slug>"
}
}
}
}
SSE — deprecated
https://mcp.plane.so/sse is maintained for backward compatibility only. Use an
HTTP transport instead.
Tools
The server advertises 30 tools, one per resource. Each takes an action
parameter that selects the operation:
workitem(action="create", project_id=..., name="Fix login")
workitem(action="list", project_id=..., pql='state__group = "started"')
cycle(action="archive", project_id=..., cycle_id=...)
Every tool's description lists its actions with their required and optional parameters, so the catalogue is self-documenting at call time.
→ Full tool and action reference
Querying work items
List, count and search accept PQL, Plane's query language:
workitem(action="list", project_id=..., pql='state__group = "started" AND priority = "urgent"')
workitem(action="count", pql='assignees__id = "<member id>"', group_by="state_id")
Call get_pql_reference for the full syntax, operators and worked examples.
Upgrading from the per-operation tools
Earlier releases exposed one tool per API operation. Existing integrations keep
working: 169 of those 177 names still resolve to the consolidated tool, so a
saved prompt or script calling create_work_item or list_cycles needs no
change. They are no longer advertised, and they keep the parameter names they
shipped with (work_item_id, not workitem_id).
Seven names chose between two operations with a parameter
(manage_project_archive(archive=False)), which one tool-and-action pair cannot
reproduce; calling one tells you its replacement. get_pql_reference is
unchanged.
Configuration
Authentication
| Variable | Required for | Purpose |
|---|---|---|
PLANE_API_KEY | stdio | API key |
PLANE_WORKSPACE_SLUG | stdio | Target workspace |
PLANE_BASE_URL | optional | Plane API URL (default https://api.plane.so) |
The remote transports carry credentials in the connection — the OAuth flow or the PAT headers — and need none of these.
Outbound proxy for attachment uploads
HTTP_PROXY / HTTPS_PROXY / NO_PROXY (and their lowercase forms) are honored
for the attachment-upload-from-URL fetch (upload_from_url). When a proxy
applies to a given request, the proxy — not this server — is that request's
SSRF control: the server does not pin or validate the target's resolved
address itself, since the proxy connects to and resolves the target
independently. Scope NO_PROXY and proxy egress policy accordingly.
Self-hosting the server itself:
| Variable | Purpose |
|---|---|
PLANE_INTERNAL_BASE_URL | Internal URL for server-to-server calls, preferred over PLANE_BASE_URL |
REDIS_URL | OAuth token storage as one connection URL (redis:// or rediss:// for TLS); wins over host/port |
REDIS_HOST / REDIS_PORT | OAuth token storage; falls back to in-memory |
PLANE_OAUTH_PROVIDER_* | OAuth client credentials and base URL |
MCP_PATH_PREFIX | Path prefix for the HTTP routes, when mounted behind a proxy — /plane serves /plane/http/mcp |
OAuth redirect URIs
The OAuth transports validate each client's redirect URI against an allowlist. Common clients (Cursor, VS Code, Claude.ai, ChatGPT connectors, localhost) are allowed by default.
To onboard a new client without a release, append patterns:
export PLANE_OAUTH_ALLOWED_REDIRECT_URIS="https://newclient.com/cb,https://other.app/oauth/*"
* matches any port, path segment or subdomain. Keep the host pinned and
wildcard only the port or path.
Logging
Structured JSON. Each tool call logs its name, duration, status and — when available — an opaque user id and the workspace slug.
export LOG_USER_INFO=false # also log the display name (PII);
export LOG_PAYLOADS=false # keep request payloads out of logs; default true
Only the OAuth and PAT transports carry a display name; stdio is unaffected.
Development
git clone https://github.com/makeplane/plane-mcp-server
cd plane-mcp-server
uv pip install -e ".[dev]"
Run the server against a workspace:
PLANE_API_KEY=... PLANE_WORKSPACE_SLUG=... python -m plane_mcp stdio
python -m plane_mcp http # port 8211
Tests, format, lint:
pytest # no network or credentials needed
ruff format plane_mcp/ tests/ # line length 120
ruff check plane_mcp/ tests/ # rules E, F, I, UP, B
The suite runs fully offline — every action of every resource is
executed against a stand-in that binds each call against the genuine plane-sdk
signature. See plane_mcp/tools/README.md.
Live integration tests are skipped unless you point them at a running server:
export PLANE_TEST_API_KEY=... PLANE_TEST_WORKSPACE_SLUG=...
export PLANE_TEST_MCP_URL=http://localhost:8211 # optional; this is the default
pytest tests/test_integration.py -v
They write real data to that workspace.
Repository layout
| Path | Contents |
|---|---|
plane_mcp/__main__.py | entry point; picks the transport from argv[1] |
plane_mcp/server.py | one factory per transport |
plane_mcp/client.py | resolves credentials into a plane-sdk client |
plane_mcp/auth/ | OAuth provider and header auth |
plane_mcp/tools/ | the tool surface: one module per Plane resource |
plane_mcp/toolkit/ | shared building blocks for the tool surface |
plane_mcp/pql_reference.py | PQL syntax reference served to models |
Contributing
Pull requests welcome. Please run pytest and ruff check before submitting; new
tools should come with the invariants described in
plane_mcp/tools/README.md.
See CONTRIBUTING.md and CODE_OF_CONDUCT.md.
Migrating from the Node.js server
@makeplane/plane-mcp-server (Node.js) is deprecated and unmaintained. This
Python implementation replaces it.
| Node.js | Python |
|---|---|
PLANE_API_KEY | PLANE_API_KEY |
PLANE_API_HOST_URL | PLANE_BASE_URL |
PLANE_WORKSPACE_SLUG | PLANE_WORKSPACE_SLUG |
Replace the command and args with the stdio configuration in
Quick start.
License
MIT — see LICENSE.
Ressources associées

BMad Method
SkillAgile AI-driven development method: analyst, product manager, UX designer and other role skills turn ideas into product briefs, PRDs and PRFAQs, then split them into tasks for agents. Needs a skills-capable agent such as Claude Code, plus Node.js, Git and uv; then run bmad setup.
bmad-code-org/BMAD-METHOD54.2k
Linear (Codex plugin)
PluginLinear's official plugin in OpenAI's Codex marketplace: search, create and update Linear issues, projects and initiatives from Codex, and draft PRDs, status updates and customer request summaries. Codex only; needs a Linear account and authorization.
openai/plugins7.5k
Task Master
MCPParses a PRD into a dependency-aware task list so Cursor, Claude Code and other coding agents break it down, implement it and track progress; also works as a CLI. Needs at least one model API key, or reuse a Claude Code or Codex CLI login.
eyaltoledano/claude-task-master28.2k
Spec Workflow MCP
MCPSpec-driven development MCP: the agent writes requirements, design and task docs in three steps, then implements task by task. A web dashboard or VS Code extension lets PMs review, comment, approve and track progress. Needs Node.js; launch the dashboard separately via npx.
Pimzino/spec-workflow-mcp4.3k
Task Master
CLIParses a PRD into a dependency-aware task list so Claude Code, Cursor and other agents pick up the next task in order, split subtasks and update progress. Offers both a CLI and an MCP server. Needs at least one model API key, or reuse a Claude Code or Codex login.
eyaltoledano/claude-task-master28.2k
Taskmaster
PluginParses a PRD into a dependency-aware task list with complexity analysis, task breakdown and research; the plugin bundles slash commands, subagents and the MCP server. Needs a model API key, or reuse a Claude Code or Codex CLI login.
eyaltoledano/claude-task-master28.2k