Overview
MCP is how an AI client reaches a workspace without you writing an integration. Taskade ships two servers, and the one you want depends on whether you are pointing a desktop client at your own workspace or running against the hosted endpoint.
Choosing a server
| Workspace MCP | Hosted MCP | |
|---|---|---|
| Where it runs | Your machine, via npx |
Taskade infrastructure |
| Endpoint | local stdio | https://www.taskade.com/mcp |
| Transport | stdio | Streamable HTTP (no SSE) |
| Auth | Personal access token | PAT, or OAuth 2.0 with the mcp scope |
| Best for | Desktop AI clients you control | Remote clients, shared setups |
| Can edit Genesis app source | No | Yes |
Authentication
The hosted server implements the full OAuth discovery chain, so a compliant client configures itself:
WWW-Authenticate: Bearer realm="mcp",
resource_metadata="https://www.taskade.com/mcp/.well-known/oauth-protected-resource"
Access tokens live 3600 seconds; refresh tokens rotate and revoke the pair they replace. A personal access token also works and skips the scope check entirely.
MCP is plan-gated per tool, not at the transport. A Free account authenticates successfully and then sees zero spaces — which reads like a bug and is not one. Any paid plan resolves it.
Direction matters
| Direction | What it means | Status |
|---|---|---|
| Inbound — Taskade as an MCP server | Claude/Cursor reach into your workspace | Live, any paid plan |
| Outbound — a Taskade automation as an MCP client | A flow step lists and calls a remote MCP server's tools | Live via the MCP Client connector |
| Outbound — a Taskade agent as an MCP client | An agent calls an external MCP server by itself | Not available on any tier |
Outbound MCP exists, but it lives in the automation engine, not the agent runtime. Point the MCP Client connector at any https MCP server and call its tools as flow steps; to give an agent the same reach, have it trigger that automation instead of calling the server directly. There is no per-agent connector toolbox and no Space Connectors screen in production, so don't build against one.