MCP deep dive
MCP is the adapter layer between an AI client and Hyperlake’s governed capabilities. The client asks for available tools and calls a selected tool with structured arguments. Hyperlake validates the arguments, resolves the target, applies policy, executes the operation through the approved path, and returns structured evidence.
Tool categories
Section titled “Tool categories”The exact list is versioned with the CLI, but the public categories are:
| Category | Examples |
|---|---|
| Identity and discovery | current user, clusters, catalogs, namespaces |
| Data | schemas, tables, read-only SQL, semantic model requests |
| Security | grants, sync status, policy tests, credential references |
| Operations | observability, monitors, diagnostics, approved services |
| Applications | approved app definitions, previews, deployments, actions |
| Agent context | agent manifests, project/cluster context, conversations |
The model receives tool descriptions and schemas, not implementation secrets. It should ask for confirmation before destructive or approval-gated actions.