Connecting the MCP
Point an AI employee at your own MCP server.
What you see
AI Workforce → Connections is where a workspace registers an outbound MCP server: a name, the server URL and an optional auth token. CoreX connects, lists the server's tools, and caches the catalog; each employee's setup screen then shows those tools alongside its built-in ones, so you grant them per employee.
This is the outbound half of MCP. The inbound half — letting any MCP client query your CoreX metrics read-only — is CoreX as an MCP server, and the two are configured independently.
Use it
- The URL you supply is checked on every connect, not just the first, so a server that later resolves somewhere unsafe is refused rather than trusted on yesterday's answer.
- The auth token is encrypted at rest and never rendered back to the browser.
- Everything a remote server returns is treated as untrusted input and fenced before it reaches the model — a tool result cannot issue instructions.
- Refresh the tool catalog after you add tools on the server side; a tool that appears mid-conversation is picked up on the next refresh.
Why it matters
Scoped, per-employee connections keep agent actions controlled — an employee can only call what you granted it, from a server you registered.