Connect an external AI client with MCP
On this page
Cloudgate's MCP connection exposes supported platform tools to an external AI client. Use the Hub's current connection instructions so the client targets the correct server and account.
Set up the connection#
Open the external AI/MCP setup from Kick Start → Apps & Tools or the available connection guidance in the Hub. Use the endpoint and authentication flow shown for your client.
The workflow MCP service uses /mcp/workflow with the supported OAuth or bearer-authentication flow. Use your deployment's actual base URL; the Hub's web address and API host may differ.
Complete sign-in in the supported flow, then verify the workspace and available tools with a read-only request such as listing controllers. Do not paste a token into a conversation as a configuration step.
Understand tool vocabulary#
| Hub label | Common MCP name |
|---|---|
| Controller | Project |
| Action / workflow API | Endpoint |
| Action Logs | Per-request execution records |
| Key Selectors | Selectors / keys |
Ask the client to inspect the execution model and node catalog before designing a graph. Existing cookbook recipes and templates can provide a better starting point than inventing node configuration.
Make edits reviewable#
Have the client inspect the current workflow and open its draft before changing it. Keep a complete export when using graph replacement operations; a slim summary may omit scripts, SQL, and request bodies.
Review the proposed changes, validate the graph, and test the sandbox route. Tool availability differs between connections, so confirm that the client actually has the required editing and recovery capabilities.
Troubleshoot access#
If a tool is unavailable or forbidden, check the account, connection scope, and platform permission. Reconnect through the supported flow when a session expires. Avoid changing a workflow's API security to solve a tool-connection problem; those are separate access boundaries.
Need to investigate a problem? Start with troubleshooting