Looker MCP and Agentic Analytics: Semantic Layer First

In a little over a year, Looker went from a BI platform with a chat feature to the semantic layer that Google positions under every analytics agent: a Model Context Protocol server, Conversational Analytics in general availability, dashboard agents, a LookML agent in VS Code. The pitch is that any assistant, Gemini or Claude, can now query your data through the model. The catch is in the last three words. The agent is only as good as the LookML beneath it. This guide summarises what has shipped, why most agentic projects still get cancelled, and the six things to fix in the model before opening it to an agent.

What has shipped, in order

Date	Announcement	What it means
9 August 2025	Looker MCP Server, built on the MCP Toolbox	Gemini CLI, Claude Desktop and other MCP clients can query Looker Explores through the LookML model
14 November 2025	Conversational Analytics generally available in Looker	Natural-language questions over one or more Explores, with the generated query and the calculation explained
22 April 2026, Next '26	Looker-managed MCP server in preview, Dashboard Agents, Visualization Assistant GA, a VS Code extension and a LookML agent	The server is now hosted by Looker; agents live inside dashboards; LookML itself gets an assistant
2 July 2026	Google cites PayPal: more than 3,000 Conversational Analytics users through Claude Desktop and Looker MCP	The first large deployment where the assistant is not Google's own

Two things stand out. The first is that Google is not tying the agent to Gemini: the PayPal case runs on Claude Desktop, through the MCP server. The second is that every one of these features goes through the same door, the LookML model. Whatever the assistant, it selects fields and filters in an Explore, and Looker writes the SQL.

What Looker MCP actually does

An MCP server exposes tools to an assistant. Looker's exposes the model: list Explores, list their fields with labels and descriptions, run a query on chosen dimensions and measures, fetch a dashboard or a Look. The assistant never sees a table name or writes a join. It reads your descriptions to decide which fields answer the question, runs the query with the user's own permissions, and formats the result.

This is the strength of the approach and its whole dependency. An LLM writing SQL against a warehouse has to guess which of the several revenue columns is the right one and how to join orders to customers without doubling rows. Through Looker, it chooses among named, described, already joined fields. If those fields are ambiguous, duplicated or undocumented, it still chooses, and it chooses with confidence.

Why most agentic projects get cancelled

Gartner predicted on 25 June 2025 that more than 40 percent of agentic AI projects would be cancelled by the end of 2027, citing rising costs, unclear business value and inadequate risk controls. In analytics, the concrete version of that prediction is easy to picture. A pilot is launched on a model that has three measures called "Revenue", hundreds of fields per Explore and no descriptions. The agent answers quickly and often wrongly. Users check a few answers, find a discrepancy, and stop asking. The project is later closed with the conclusion that "AI does not work on our data".

The failure is rarely in the agent. It is in the semantic layer the agent was pointed at, and it was visible before the pilot started. This is why the order matters: audit and fix the model first, then connect an assistant.

Six prerequisites on the LookML side

These are the checks that decide whether an agent, through MCP or Conversational Analytics, will answer correctly.

1. One measure per metric

No duplicated measures across views, no dead fields with similar names. If the model contains two definitions of margin, the agent will alternate between them and nobody will know which one produced a given answer. The LookML audit checklist covers how to find and merge them.

2. Descriptions that are current

The agent reads label and description to choose a field. A description that says what the measure includes and excludes, its reference date and the synonyms users actually say is the single highest-return change you can make. A description that is out of date is worse than none: it steers the agent into a confident mistake.

3. Technical fields hidden

Surrogate keys, ETL timestamps, raw JSON columns and helper dimensions should carry hidden: yes. An agent presented with a field called amount_raw next to net_revenue may pick either.

4. Explores limited to what they answer

One Explore per business domain, with fields: restricting the list to what users pick. A narrow Explore gives better answers than a wide one, for the same reason a short menu is easier to order from. Conversational Analytics works on one to five Explores per agent; keep them tight.

5. Aggregate awareness for cost

Agents generate many more queries than dashboards do, because every question is a new query and users iterate. Without aggregate tables and sensible caching, a successful pilot becomes a BigQuery cost problem within weeks. Set up aggregate awareness on the Explores you expose before the traffic arrives, not after.

6. Fifty reference questions with expected answers

Collect fifty real questions from the future users and compute the expected answer by hand, from the current dashboards or from SQL. This set does three jobs: it measures the agent before anyone else uses it, it feeds the verified queries that improve nearby answers, and it makes the go or no-go decision at the end of the pilot a number rather than an impression.

FAQ

Does Looker MCP work with Claude, or only with Gemini?

With both. The MCP server was announced with Gemini CLI and Claude Desktop as clients, and the PayPal deployment Google cites in July 2026 runs on Claude Desktop.

Can the agent bypass row-level security?

Not through Looker. Queries run with the authenticated user's roles, access grants and access_filter. Which is exactly why security has to live in the model rather than in hand-filtered dashboards: anything outside the model does not exist for the agent.

Next step

Neuravoid runs the agentic analytics pilot as a fixed-price, four-week engagement: the model audited and fixed on one domain, one agent built and scored on your reference questions, MCP access opened to the assistant your team uses, and a measured decision at the end.

See the agentic analytics pilot, and read next: Conversational Analytics in Looker: 6 prerequisites and Why your revenue numbers never agree.

Sources

• Introducing Looker MCP Server, Google Cloud Blog, 9 August 2025
• Looker Conversational Analytics now GA, Google Cloud Blog, 14 November 2025
• Looker updates for agentic BI at Next '26, Google Cloud Blog, 22 April 2026
• Looker in the 2026 Gartner Magic Quadrant, Google Cloud Blog, 2 July 2026
• Gartner predicts over 40 percent of agentic AI projects will be canceled by end of 2027, Gartner, 25 June 2025