Looker MCP

What does a Looker MCP server do?

A Looker MCP server lets an AI assistant query your LookML models and, where an admin allows it, build content and edit LookML. Looker generates the SQL from its semantic layer, so the assistant doesn't write its own. Looker sends uncached queries to your database, so those calls use warehouse compute.

• Looker: Google Cloud's BI platform, with LookML models, Explores, Looks and dashboards. Google's tool list covers these and names no Looker Studio reports.
• Looker Studio: a separate reporting product, now called Data Studio. Google hasn't published an MCP server for it, and Data Studio's release notes don't mention MCP.
• Data Studio can read Looker data through its own BI connector, and those API calls aren't metered. That connector is separate from MCP.

Is there an official Looker MCP server?

Yes, and Google publishes two. The Looker-managed MCP server is built into Looker-hosted instances and reached general availability on 26 October 2026. MCP Toolbox for Databases is an open-source server you run yourself. Google says Toolbox isn't a supported Google Cloud product, and it recommends the managed server for Looker-hosted instances.

• Looker-managed MCP server: available on Looker-hosted Looker (original) and Looker (Google Cloud core). Customer-hosted Looker (original) can't use it.
• MCP Toolbox for Databases: Apache-2.0, in Google's googleapis/mcp-toolbox repo, renamed from genai-toolbox. Google sends support questions to the repo's GitHub Issues page.

What is the Looker MCP server URL, and how do you sign in?

The managed server URL is LOOKER_INSTANCE_URL/mcp, so it differs per instance. You sign in through OAuth 2.1 with your normal Looker login, and the assistant inherits your roles. A Looker admin must enable your client in Manage OAuth Clients first. Toolbox signs in with a Looker API client ID and secret, or OAuth.

• The Application Library has ready-made clients for Cursor, VS Code, Claude Code, Claude Desktop, ChatGPT, Gemini CLI, Cline, Windsurf and Antigravity. An admin turns one on and copies its client ID.
• Any other tool needs a custom client with exact redirect URIs, because Looker doesn't allow wildcards. An admin can restrict a custom client to one group.
• Toolbox needs an HTTPS reverse proxy, such as Cloud Run, for OAuth. By default it listens on 127.0.0.1, port 5000.

What can a Looker MCP server read and change?

That depends on what a Looker admin enables and what your own role allows. All tools start disabled, and the server enforces user-level permissions. Reads cover models, Explores, Looks, dashboards and instance health. Writes create Looks and dashboards and edit LookML files, so a read-only setup leaves the write tools off.

• The MCP admin page shows read or write beside each tool, under managed names such as looker_query and looker_update_project_file. Google recommends enabling only what each use case needs.
• The managed server page links to the Toolbox tool list below, where names have no looker_ prefix. Google hasn't published a separate list for the managed server.
• On Toolbox, the looker config leaves out LookML file tools, which sit in the looker-dev config. The looker config still includes make_look and make_dashboard.

What it can read

• Models and fields: get_models, get_explores, get_dimensions, get_measures, get_filters and get_parameters.
• Queries: query returns data, query_sql returns the SQL Looker generates and query_url returns a Looker link.
• Saved content: get_looks, run_look and get_dashboards.
• Instance health: health_pulse, health_analyze and health_vacuum.
• LookML and connections: get_projects, get_project_file, get_connections, get_connection_tables and get_connection_table_columns.

What it can change

• Saved content: make_look, make_dashboard and add_dashboard_element create Looks, dashboards and tiles.
• LookML: create_project_file, update_project_file and delete_project_file change model files. Google requires Development Mode for LookML authoring, and dev_mode turns it on for the session.
• Toolbox's looker-dev config adds create_view_from_table, directory tools and Git branch tools. It lists no commit or deploy tool.

How do you set up a Looker MCP server in Claude, ChatGPT and Cursor?

Start in Looker. An admin enables the tools you need, then turns on your client in Manage OAuth Clients and copies its client ID. Add LOOKER_INSTANCE_URL/mcp to your client and sign in. Google documents Claude Desktop, Claude Code and Cursor. It lists ChatGPT as a client but gives no steps.

{
  "mcpServers": {
    "looker": {
      "url": "LOOKER_INSTANCE_URL/mcp",
      "auth": {
        "CLIENT_ID": "CLIENT_GUID"
      }
    }
  }
}

What can you ask a Looker MCP server?

Good first prompts read the model and ask for the SQL before anything runs, so you can check the logic. These examples fit the tools Google lists, so swap in your own model and Explore names. The last one needs the dashboard write tools on, and it saves content in Looker.

• Reads

  List the Explores in our marketing model, then list the measures in the one that holds campaign spend. Don't run a query yet.

  Copy

• Reads

  Using that Explore, show the SQL Looker generates for spend and conversions by channel over the last 30 days. Don't run it.

  Copy

• Reads

  Run that query with a 500 row limit and give me the Looker link so I can check it.

  Copy

• Reads

  Find LookML elements in this project that might be unused. List them and don't change anything.

  Copy

• Writes

  Create a dashboard called Weekly Paid Media with that query as a tile and give me the URL. Show me the plan first.

  Copy

What are the limits of a Looker MCP server?

Tool calls count toward your Looker API limits, and OAuth scopes aren't supported yet. Google's server page still words two limits as preview limits: no Dynamic Client Registration, and fixed capacity that can cause timeouts at peak. Toolbox warns that its prebuilt tools are pre-1.0 and can change between versions.

• Monthly API limits on Looker (Google Cloud core) are 1,000 query-based and 1,000 administrative calls on Standard, 100,000 and 10,000 on Enterprise, and 500,000 and 100,000 on Embed. Google doesn't say what happens at the limit.
• Tool changes aren't pushed to connected clients, so wait 30 seconds and reconnect. Calls to a disabled tool are rejected even if a client still shows it.
• Toolbox's query tool defaults to 500 rows, and -1 means unlimited, which can mean a large warehouse query.
• Disabling a client revokes its refresh tokens, and access tokens already issued expire within 60 minutes.
• Agent activity is logged in System Activity, and also in Cloud Audit Logs on Looker (Google Cloud core).

How much does a Looker MCP server cost?

Google says the managed server is available at no additional cost. You still pay for Looker, and Looker (Google Cloud core) is sold by edition on annual contracts. Your warehouse bills the queries Looker sends it, and tool calls use your API quota. Toolbox is open source, so you pay for wherever you host it.

• Edition details sit on Google's Looker pricing page.
• Looker reuses cached results when a query matches exactly, which saves a warehouse run.

What are the alternatives to a Looker MCP server?

The main alternatives are a warehouse server, dbt's semantic layer, or another BI tool's server. For Looker Studio reports, connect the data underneath them, because Google hasn't published a Data Studio server. Google's BigQuery server is one route, and our Looker Studio implementation page covers the reports themselves.

• BigQuery MCP: Direct warehouse queries from Google's server, without LookML definitions.
• dbt MCP Server: Semantic Layer metrics, if your definitions live in dbt.
• Tableau MCP: Another BI tool's route, if you run Tableau beside Looker.
• Power BI MCP: Another BI tool's route, if you run Power BI beside Looker.

Which skill goes with a Looker MCP server?

Use the free Agent write safety skill before you enable any write tool, because a LookML edit can change the numbers every dashboard shows. Use the Marketing context kit to write down your model names and metric definitions once, so every prompt starts from the same rules and the assistant has less to guess.

• Agent write safety: Sets approvals before an agent can save content or edit LookML.
• Marketing context kit: Holds your model names and metric definitions in one folder for every prompt.