An MCP turns a hand-picked set of your FlowFn capabilities into a Model Context Protocol server that external AI clients — Claude Desktop, Cursor, the MCP Inspector, and anything else that speaks MCP — can connect to and use. It's a top-level, App-owned artifact, alongside Workflows, Agents, and Streams.
This is FlowFn hosting an MCP server. It's the opposite direction of an agent connecting out to an external MCP server — here, outside AI clients connect in to your tools.
MCP hosting is available on Pro plans and above. On other plans the MCP page shows an upgrade prompt; anything you built before a plan change stays intact and unlocks again on upgrade.
Create an MCP
- Open MCP in the dashboard and choose New MCP.
- Pick the App it belongs to and give it a name.
- Save. You now have an empty MCP with its own connection URL and token.
Attach what it exposes
In the editor, attach the capabilities you want external clients to be able to list and call. A client can use exactly what you attached — nothing else. Four kinds:
- Platform-tool actions — e.g. a Slack "post message" action. (See eligibility below — sensitive tools are off by default.)
- Data sheets — expose a sheet for read (query rows) or read & write (create / edit / delete). Password/secured columns are never returned.
- Workflows — the client can trigger a run.
- Agents — the client can run the agent.
Exposing a workflow or agent is high-privilege: an external client can then trigger everything that workflow or agent does — including sending email, charging payments, or writing data. The editor flags this. Only attach ones you're comfortable letting an outside client run.
Safe by default (eligibility)
Sensitive platform tools (payments, cloud, raw databases, and similar) do not appear in the attach picker unless a FlowFn admin has explicitly enabled them for the MCP surface. This deny-by-default posture mirrors the interactive-agent tool picker — nothing sensitive is exposed to an external endpoint by accident. If a tool you expect is missing from the picker, ask an admin to enable it for MCP hosting.
Connect a client
The editor shows your connection URL (e.g. https://mcp.flowfn.com/mcp/v1/<code>) and a single rotatable bearer token. Point any MCP client at the URL and send the token in the Authorization header. The editor also gives you a copy-ready client-config snippet you paste straight into your client (Claude Desktop / Cursor), so connecting is copy-and-go.
Rotate the token any time from the editor — the old token stops working immediately, on the very next request, for every connected client. Disable the MCP to reject all clients until you re-enable it.
Limits & credits
Each MCP tool call is metered against a per-day limit and charges a small per-call credit fee to the account owner (separate from any credits a triggered workflow or agent run consumes on its own). When the account runs out of credits, calls are refused until it's topped up. Your plan sets how many MCP servers you can host and the daily call cap.