How to Connect Metabase to Claude Using MCP

What Changed in Metabase 60

Before April 2026, connecting an AI assistant to Metabase required community-built packages that needed your Metabase username, password, or a manually created API key stored in a config file. That setup worked but broke whenever Metabase updated its session API.

Metabase 60 shipped an official Model Context Protocol (MCP) server built into the platform itself. MCP is an open standard that AI clients like Claude, Cursor, and VS Code Copilot use to call external tools. The Metabase implementation means the platform becomes a first-class data tool that any MCP-aware AI assistant can query — without third-party packages or stored credentials.

As of launch, Metabase's MCP server supports Anthropic models only. You supply your own API key and pay directly for token usage.

What You Need Before Starting

• Metabase 60 or later (self-hosted or Metabase Cloud)
• Admin access to your Metabase instance
• An Anthropic API key (create one at console.anthropic.com)
• An MCP-compatible AI client: Claude Desktop, Claude Code, Cursor, or VS Code with the MCP extension

Step 1: Enable AI in Metabase Admin

Log in to your Metabase instance as an admin. Go to Admin > AI from the top navigation.

On the AI settings page:

1. Select Anthropic as your provider from the dropdown.
2. Paste your Anthropic API key into the API key field.
3. Click Connect.
4. Once connected, choose a model from the Model dropdown (Claude Sonnet is the default recommendation).

After you save, Metabase enables four features: Metabot (natural-language question answering), inline SQL generation, the MCP server, and the Agent API. Each has its own toggle if you want to enable them selectively.

Step 2: Enable the MCP Server Toggle

Still in Admin > AI, find the MCP Server section and flip the toggle to on. This activates the endpoint at:

https://your-metabase.example.com/api/mcp

The server uses OAuth 2.0 for authentication. Metabase runs its own embedded OAuth server, so you do not need to configure an external identity provider. When a user connects their AI client, they complete a short browser-based OAuth flow and the client stores a scoped token. All MCP queries run under that user's Metabase permissions — row-level security and collection access policies apply automatically.

Step 3: Connect Claude Desktop

Open your Claude Desktop config file:

• macOS: ~/Library/Application Support/Claude/claude_desktop_config.json
• Windows: %APPDATA%\Claude\claude_desktop_config.json

Add or update the mcpServers section:

{
  "mcpServers": {
    "metabase": {
      "url": "https://your-metabase.example.com/api/mcp",
      "transport": "streamable-http"
    }
  }
}

Save the file and restart Claude Desktop. On first use, Claude will open a browser tab for the Metabase OAuth flow. Approve the connection and you're done.

Step 3 (Alternate): Connect Claude Code

If you use Claude Code in the terminal, run:

/mcp add metabase https://your-metabase.example.com/api/mcp --transport streamable-http

Claude Code handles the OAuth redirect automatically.

Step 4: What You Can Do Once Connected

The MCP server exposes five core tools to the AI client:

• search — find tables, metrics, and saved questions by keyword
• get_table — retrieve schema details for a specific table
• get_table_field_values — sample values from a column (useful for filter suggestions)
• construct_query — build a structured Metabase query from a description
• execute_query — run a query and return results with column metadata

In practice, you can open Claude and ask: "What were total orders by region last quarter?" Claude calls search to locate the relevant table, construct_query to build the filter, and execute_query to return results — all in one message. The query runs under your permissions, so you only see data your Metabase role allows.

Step 5: Verify the Connection

After connecting your AI client, run a simple test. In Claude, type:

"List the tables available in my Metabase."

If the MCP connection is working, Claude will call the search tool and return a list of tables from your instance. If it fails, the most common causes are:

• The MCP server toggle is off in Admin > AI.
• The instance URL in your config has a trailing slash (remove it).
• Your Anthropic key has expired or hit its rate limit.

You can also test the endpoint directly with curl:

curl -I https://your-metabase.example.com/api/mcp

A 401 response confirms the endpoint is live and waiting for OAuth. A 404 means the MCP toggle is off.

Permissions and Security Notes

Every query the AI client runs is scoped to the authenticated user's Metabase permissions. An analyst with read-only access to three collections cannot use the MCP connection to query tables outside those collections. This makes the setup safe for teams with multi-tenant data or strict row-level security.

The Disable all AI features toggle in Admin > AI is a master kill switch. Turning it on hides Metabot, inline SQL, the MCP server, and the Agent API simultaneously. Use it if you need to temporarily pause AI access without removing your API key configuration.

Common Issues

OAuth redirect fails on self-hosted instances: Set the MB_SITE_URL environment variable to the public URL users access Metabase from. Without it, the OAuth callback URL will point to localhost.

Model not available: Metabase 60 only supports Anthropic. If your team uses another provider, you will need to wait for future Metabase releases or use a community MCP package that accepts any OpenAI-compatible endpoint.

Slow query responses: The MCP server calls Metabase's internal query engine, which is subject to the same timeouts as the UI. Long-running queries that time out in the dashboard will also time out in the AI client.

Summary

Metabase 60 makes the AI-to-analytics connection straightforward: add an Anthropic key in admin settings, enable the MCP toggle, and point your AI client at /api/mcp. The OAuth flow handles credentials, and existing Metabase permissions carry over automatically. For teams already running Metabase as their BI layer, this is the lowest-friction path to querying data in plain English without rebuilding any infrastructure.