Add MCP server to expose catalog as AI-agent tools

[ravisuhag]

Context

As part of Compass's evolution into an AI serving layer, we need to expose the existing catalog functionality as MCP (Model Context Protocol) tools. This is the fastest path to making Compass usable by AI agents — it's a thin adapter layer over existing APIs with no storage changes required.

Scope

Wrap existing Connect RPC endpoints as MCP tools:

• search_assets — semantic search across the catalog
• get_lineage — traverse the graph from any starting node
• get_schema — full schema with column descriptions and types
• get_owners — who to ask about this asset
• get_context — composed context document (depends on context composition endpoint)

Approach

• Implement an MCP server mode (compass serve --mcp or built into the existing server)
• Each tool maps to existing Compass service methods
• Response formats optimized for LLM consumption (structured text, not raw JSON)
• Authentication should carry through from existing auth mechanisms

Why

AI agents, copilots, and autonomous systems are the fastest-growing consumers of organizational metadata. MCP is the emerging standard for tool-based AI integration. Shipping this immediately validates the new direction with zero changes to the data layer.

References

• Compass Roadmap — AI Serving Layer
• Model Context Protocol Specification