Collibra's Model Context Protocol server connects Snowflake Cortex Agents (CoWork and CoCo) to governed metadata, enabling agents to discover data assets, trace lineage, look up business glossary terms, and manage data contracts using natural language, all while respecting per-user OAuth-based role access controls from Collibra. A financial services use case shows a risk analyst using Cortex Agents to verify data certification, trace measure definitions, inspect lineage, and update catalog attributes without leaving Snowflake. The integration is framed as an open, interoperable layer (also usable with Claude, Databricks, and other MCP-compatible platforms) that keeps governance controls intact as AI agents take on more work.
Table of contents
What Is Collibra MCP?Connecting Snowflake Cortex to Collibra MCPSecurity by design: Per-user OAuth identityThe MCP tool paletteUse Case: Financial Services — BCBS 239 regulatory capital reportingGet Anand Akella’s stories in your inboxWhy this integration mattersGetting startedQuestions this post answers
What is the Collibra MCP server and what does it let AI agents do in Snowflake?
The Collibra MCP server is a Model Context Protocol implementation that gives AI assistants a live, structured connection to Collibra's governed metadata and business context. It exposes read tools (like discovering data assets and business glossary terms), write tools (creating or editing catalog entries), and experimental opt-in tools, using streamable HTTP transport and OAuth authentication for enterprise security. Teams wiring governed data catalogs into AI agents can follow integration patterns like this one on daily.dev.
How does per-user access control work when Cortex Agents query Collibra through MCP?
Each user authenticates with their own Collibra credentials via OAuth 2.0, so every API call from a Cortex Agent carries that user's authenticated identity. Collibra's role-based access controls apply exactly as they would for manual access, meaning a user without permission to view a sensitive asset cannot retrieve it through the agent either, keeping the workflow audit-ready without a separate identity layer. Anyone designing secure AI agent access to enterprise data can compare governance approaches on daily.dev.