MCP servers

The MCP gateway lets you give AI agents governed access to tools. You register remote Model Context Protocol (MCP) servers per project; the gateway then fronts them with the same identity, isolation, and audit you use for LLM traffic.

Who can do this

Org admins (for their organization) and platform admins, on Projects → MCP Servers.

How a governed MCP request flows

A request must carry a valid project API key (key-auth) and may only reach its own project's path (tenant guard) — a key from another project gets 403.

Register a server

1. Open Projects → MCP Servers and click Add server.
2. Enter a name, the backend MCP server URL, a timeout, and (if the backend needs one) an upstream credential.
3. Toggle enabled and save. The control plane provisions the route and isolation rules.
4. Share the generated connect URL with your developers: https://mcp.<your-domain>/<organization>.<project>/<server-name>.

What's governed (Stage 1)

• Access — agents authenticate with the project API key (one key for chat and tools).
• Isolation — strict per-project: a key cannot reach another project's servers.
• Activity — tool-call activity is recorded per organization.

Scope

This first stage covers governed access + a catalog of remote MCP servers. Per-tool budgets/guardrails and auto-generating MCP servers from REST APIs are future capabilities.

Next steps

Developers connect to these servers via Use MCP servers.