Looker AI: What It Is & How To Use It In 2026

"Looker AI" means three different Google products in 2026. Gemini in Looker is a set of assistants inside Looker for charts, expressions, LookML and Python analysis. Conversational Analytics lets people ask questions of Looker Explores in plain language, and has been generally available since November 2025. The Looker MCP server, either the open-source MCP Toolbox or the newer Looker-managed server, lets an outside AI agent such as Claude Code, ChatGPT or Gemini CLI query through your LookML.

All three answer through your LookML, so their numbers match your dashboards. They differ in who can ask, where the answer shows up and how it is billed. Everything below is as of October 2026, with Google's documentation linked.

Looker AI at a glance

Gemini in Looker: the assistants inside Looker

Gemini in Looker is Google's umbrella name for the generative AI features in Looker. At Google Cloud Next on April 22, 2026, Google listed them with their status in its agentic BI announcement:

• Visualization Assistant (GA): describe a chart change in plain language and Gemini writes the chart configuration, on HighCharts-based charts such as most column, bar and line charts (Google's docs).
• Expression Assistant (preview): generates custom dimensions, measures and filters from a description of the logic.
• Insight Assistant (preview): writes summaries and highlights trends in reports.
• LookML AI agent and the Looker extension for VS Code (preview): turns a description into LookML for developers.
• Advanced Analytics, formerly the Code Interpreter: writes and runs Python inside a conversation, GA on May 7, 2026 per the Looker release notes. It needs Looker 25.8 or later on original instances and 25.10 or later on core, and is off by default even when Gemini is on.
• Agentic Workflows (preview): agents that watch metrics for irregularities and run scheduled reporting routines.

What you need to turn it on. A Looker admin enables Gemini in Looker for the instance (on Looker (Google Cloud core), in the Google Cloud console), and each user needs a role with the gemini_in_looker permission. A Looker (original) instance must be Looker-hosted and on 25.2 or later (Gemini in Looker overview).

What it costs. The Looker pricing page has no separate Gemini in Looker line; its one AI meter is Conversational Analytics data tokens (below). Google's docs note that some Gemini features were free during preview and may need an additional fee once the preview ends. For the platform editions, see our Looker pricing breakdown.

Looker Conversational Analytics

Conversational Analytics is the "chat with your data" part of Looker AI, generally available since November 14, 2025 to "all Looker platform users" once an admin enables it. It only uses fields your data team defined in LookML, and once the fields are chosen, Looker translates them to SQL deterministically (Google's overview).

What it does

• Chat with an Explore. Ask a question, get a chart or table, follow up, and ask "How was this calculated?" for a plain-language explanation of the query.
• Data agents. Analysts build agents with instructions, common questions, filters and style rules, then share them. An agent can connect to as many as five Explores; a direct Explore conversation queries one at a time.
• Verified queries. "Golden" queries an agent should reuse, GA from late August 2026 per the release notes.
• Dashboard agents. Summaries and questions inside a dashboard; the release notes list data agents for user-defined and LookML dashboards as GA on October 14, 2026.
• Row-level security. Google's July 2026 update says Looker automatically enforces row-level security during the chat.

Editions, permissions and billing

Conversational Analytics runs on both Looker (Google Cloud core) and Looker (original), on Looker 25.0 or later. The setup guide asks for a role with access_data and gemini_in_looker on each model a user queries (newer releases add chat_with_agent and chat_with_explore).

Billing is in "data tokens", pooled per Looker instance: input covers the prompt, history and context such as agent instructions, and output covers the answer, generated query, reasoning and visualization. Each platform subscription includes a monthly allowance, according to the Looker pricing page:

Looker platform tier	Included input data tokens per month	Included output data tokens per month
Standard	60M	1.2M
Enterprise / Advanced	300M	6M
Embed / Elite	1.2B	24M
Non-production and add-on instances	6M	0.12M

Allowances reset monthly and don't roll over. Overage is $3.00 per 1M input tokens and $20.00 per 1M output tokens, billed monthly. For now, Google says customers "currently enjoy promotional access without quota limits or overage fees (within fair usage limits)", that it is extending that promotion, and that it will give at least 90 days' notice before billing and quota enforcement begin. Heavy chat use on a Standard instance can outgrow its 1.2M included output tokens once it does.

Where the answers show up

Inside Looker (Explores, data agents, dashboards), in embedded apps (GA per the Next '26 post), through the Conversational Analytics API (GA for BigQuery and Looker on June 23, 2026 per its release notes), and in Gemini Enterprise, where teams can publish agents. None of the Google sources I read describe asking it from Slack or Microsoft Teams; you'd build that yourself on the API. "Looker Studio AI" is a different product: in Data Studio (Looker Studio until April 2026), business users chat with agents built in BigQuery, and Data Studio Pro adds AI features (Google Cloud blog); see our Looker Studio alternatives guide.

Known limits

• Only what's modeled: a question about a table no Explore covers, a spreadsheet or an ad platform has no answer until someone models it.
• One Explore per direct conversation, five per data agent, so questions that cross subject areas need a curated agent, and agents, instructions and verified queries need an owner, like dashboards do.
• Up to 50,000 rows per query in Looker. The API returns at most 5,000 rows of Looker data, and its known limitations warn that large results reduce accuracy.
• No maps, heatmaps or charts with tooltips; no values for filter-only fields defined as LookML parameters; verified queries only in data agents, and not with custom fields or pivots.
• Not yet inside the FedRAMP High or Medium authorization boundaries.

The Looker MCP server: two ways to get one

MCP (Model Context Protocol) is the standard way to give an AI client tools. Google's summary at Next '26: "Looker supports open-source Model Context Protocol (MCP) via MCP Toolbox, and now we're offering an all new managed MCP server that's native to Looker."

Option 1: MCP Toolbox for Databases (you run it)

Google's open-source MCP Toolbox ships a prebuilt Looker toolset, with tools such as get_explores, query, query_sql (the SQL Looker would run), run_dashboard and make_dashboard, in two prebuilt sets: looker and looker-dev, which adds developer tools.

Create API credentials (client ID and secret) for a dedicated Looker user, make sure Node.js is installed, and add the server to your MCP client's config (for example settings.json in Gemini CLI or .mcp.json in Claude Code):

{
  "mcpServers": {
    "looker": {
      "command": "npx",
      "args": ["-y", "@toolbox-sdk/server", "--prebuilt", "looker", "--stdio"],
      "env": {
        "LOOKER_BASE_URL": "https://yourcompany.looker.com",
        "LOOKER_CLIENT_ID": "your-client-id",
        "LOOKER_CLIENT_SECRET": "your-client-secret",
        "LOOKER_SHOW_HIDDEN_MODELS": "false",
        "LOOKER_SHOW_HIDDEN_EXPLORES": "false",
        "LOOKER_SHOW_HIDDEN_FIELDS": "false"
      }
    }
  }
}

Two details matter for security. The LOOKER_SHOW_HIDDEN_* settings default to true, so the agent sees what your LookML hides unless you turn them off as above. And the credentials belong to one Looker user, so everyone using that config queries with that user's access: give it a narrow role, and leave out looker-dev unless the agent should work on LookML.

Option 2: the Looker-managed MCP server

The Looker-managed MCP server, announced in preview at Next '26, embeds an MCP endpoint in the instance itself. Per the docs and admin settings page:

• It works on Looker-hosted instances. Customer-hosted Looker (original) instances should use MCP Toolbox.
• Every tool is disabled by default. An admin enables tools one by one, or uses "Enable all tools", which also turns on tools Google adds later.
• An admin registers each AI client (Cursor, VS Code, Claude Code, ChatGPT, Gemini CLI and others) as an OAuth client app.
• Each person signs in with their own Looker login, so the agent sees what that person can see. This is the main improvement over a shared API key.
• Usage shows up in System Activity Explores and Cloud Audit Logs.

The documented preview limits: OAuth scopes aren't supported yet, so access control comes down to the global tool allowlist plus each user's own permissions; Dynamic Client Registration isn't supported; and changes to the tool list aren't pushed to connected clients (users reconnect after about 30 seconds).

What an MCP server does not give you

An MCP server is a set of tools, not an analyst. The model inside each person's client picks the Explore and fields and reads the result, so answer quality depends on that model and on your LookML descriptions, and one person's corrections don't reach anyone else.

How GetDot works with Looker

GetDot is an AI data analyst that connects to Looker with an API client ID and secret and works through Looker's API. Everything below is described on our Looker integration page and in our changelog.

• Questions become Explore queries that Looker compiles. Dot writes a query against an Explore (fields, filters, sorts, pivots, a row limit), and Looker compiles it to SQL from your LookML and runs it on your Looker connection. The query view links back to the same query in Looker. Explores and fields an admin switches off in Dot can't be queried by anyone, admins included.
• Default filters respected. Since our August 2026 release, Dot sees each Explore's default filters instead of overriding them.
• Dashboards as Looker draws them. Dot can run a tile's exact saved query to verify its number, and flags tiles whose logic lives in custom filters or table calculations.
• Row-level security from Looker user attributes (Team plan). Rules can key on user-attribute values synced from Looker, so a rule such as WHERE country IN (${attr.country_code}) filters every query to that person's rows. A person with no value is denied rather than shown everything (September 2026 release).
• Read-only until an admin says otherwise. Editing dashboards and Looks, schedules and user management are each a separate permission, off by default.

Dot answers in Slack, Microsoft Teams, email and the web app, with the query shown, and can schedule an answer as a report or alert. It is also an MCP server itself, so your client asks an analyst instead of driving raw Looker tools. It also reaches the warehouse under Looker, such as BigQuery (see our BigQuery MCP guide), for questions LookML doesn't model.

Choco gave its stakeholders Dot as the front door to data in Looker and Databricks and reports 3,300+ hours saved a year and an estimated 18x ROI. At KRY, one open-ended question on the LookML behind Looker surfaced a revenue opportunity worth €800K a year, identified and sized in 10 minutes (KRY case study). These are our customers' published figures.

Dot is SOC 2 Type II and GDPR compliant, and its LLM providers run with zero data retention. Pricing (October 2026): Free with 300 one-time credits and full Pro features; Pro $180/month billed annually ($200 month-to-month); Team $720/month billed annually ($800 month-to-month); unlimited users, billed on credits, not seats.

Gemini in Looker vs Looker MCP vs GetDot

For a feature-by-feature version, see Gemini in Looker vs Dot. If you're still choosing a BI tool, our Looker vs Power BI comparison covers the platforms, and our semantic layer guide explains why LookML grounding matters for any AI on top of it.

Which one to choose

• Gemini in Looker when your users already work in Looker every day, your LookML is well described and questions fit one Explore (or five per agent), or you embed analytics on Looker Embed.
• A Looker MCP server when your analysts and engineers already work in Claude Code, Cursor or Gemini CLI and the job is exploring LookML, checking SQL or building dashboards from an IDE. Prefer the managed server's per-user OAuth where you can.
• GetDot when most questions come from people who never open Looker, when questions cross Looker and data LookML doesn't model, or when you want scheduled reports and alerts and would rather pay for usage than license more Looker seats. See how conversational analytics works end to end, or create a free GetDot account and connect Looker with an API client ID and secret.

FAQ