An agent that queries the warehouse directly skips the LookML definitions and each user's permissions, so its answers can differ from the dashboards for the same metric. Send every agent, inside Looker or outside it, through Looker's Explores and the MCP server, where each user's own permissions apply. Test the model in CI and register verified queries so the definitions those agents read stay correct.
When you need this
- Coding agents and AI tools are starting to write LookML.
- Teams want Claude, Gemini CLI, or Gemini Enterprise to query company data.
- Agents give different answers from dashboards for the same metric.
Evaluation criteria
Write LookML for agents
Can an agent tell which field a business term means?
Look forBusiness labels, descriptions that define each field, synonyms, hidden technical fields, and field-level refinements that tell the agent which field to use. Suggestion parameters keep the agent from running heavy sampling queries.
Keep logic in the model
Does business logic live where agents can see it?
Look forCommon custom fields and table calculations promoted into LookML dimensions and measures, and agent instructions that point at LookML and avoid restating definitions.
Test before merge
What stops a broken measure reaching production?
Look forLooker Continuous Integration running SQL, assert, content, and LookML validation on every pull request, plus a linter that requires a description on each field. Carousell's Google Cloud customer story describes both checks blocking merges.
Expose through MCP
How do outside agents reach Looker?
Look forThe Looker-managed MCP server on Looker-hosted instances, with tools off by default until an admin enables them. Each client registers as an OAuth client, and agents act with the signed-in user's roles.
Answer through the model
Should agents query the warehouse directly?
Look forAgents that answer through Looker's Explores and the MCP server, so LookML definitions and each user's permissions apply. An agent that queries the warehouse directly skips both.
Questions to ask vendors
- Which Explores are ready for agents, and which still depend on custom fields?
- Who reviews LookML written by AI tools before merge?
- Which MCP tools are read-only, and which can change project files?
- Do outside agents act as the user or as a service account?
Red flags
- Agents given direct warehouse access alongside Looker, with no shared definitions.
- LookML merged from AI tools without a measure test.
- All MCP tools enabled with the Enable all tools toggle.
- An API key with admin rights shared across agent clients.
Products in this piece
Sources
- Official documentation
- 7
- Public source
- 2
- Looker docs: Best practices for configuring Conversational Analytics
- Looker docs: Create and manage Explore data agents
- Looker docs: Looker-managed MCP server
- Looker docs: Admin settings, Model Context Protocol (MCP)
- Looker release notes
- Google Cloud blog: Welcome to the agentic BI era with Looker (Apr 22, 2026)
- Google Cloud customer story: Carousell
- Reddit post on the pre-GA Looker-managed MCP server and Claude's web connector (Jun 17, 2026)
- G2 review on company-wide LookML metrics and natural-language queries (Jan 15, 2026)