Metabase MCP Server

The Metabase MCP Server links your AI assistant to Metabase, a widely used open-source business intelligence platform where teams query their databases and turn the results into charts and dashboards. MCP, the Model Context Protocol, is an open standard that lets assistants such as Claude Code or Cursor work with outside services. Metabase builds the server itself. There is nothing extra to install, because it ships with the product and runs on your own Metabase instance. The release page lists it from Metabase 60, published in April 2026.

What is the Metabase MCP Server good for?

At many companies, the knowledge about key figures already lives in Metabase: curated models, agreed metrics, a glossary and dashboards. An assistant with raw database access knows none of that and often has to guess which table is the right one. This is the gap the server closes. The assistant first searches your existing content and then works from the definitions your team has settled on. You can ask things like: How did revenue per region develop last quarter? Which dashboard shows the churn rate? Turn that into a question and save it in the Marketing collection. Developers benefit in the editor as well, since they can check real numbers while building a feature without switching windows.

The tools at a glance

The documentation lists 23 tools. Reading tools include search, browse_collection, browse_data, get_content and glossary. Queries run through execute_query and run_saved_question. According to Metabase, execute_query returns 100 rows per call by default and at most 2000 per page. Writing tools such as question_write, dashboard_write, document_write, metric_write and collection_write create content, change it or move it to the trash. The documentation says these tools offer no permanent deletion of content. With alert_write and subscription_write you get alerts and subscriptions that send data by email or Slack. In clients that can display it, such as Claude Desktop in chat mode, visualize_query draws charts right inside the conversation. Claude Code does not offer this tool.

Permissions and scopes

Every tool acts on your behalf, so the assistant only sees what you may see in Metabase yourself. On top of that come scopes, which are finely graded permissions you confirm at sign-in: read content, write content, run queries, run raw SQL and create notifications. A scope never extends your own rights. Pay particular attention to execute_sql. The documentation explicitly warns that this tool can overwrite or delete existing data. It only works if you are allowed to run native queries on that database. Before connecting anything, check whether the database account stored in Metabase has write access.

Open source, Pro and Enterprise

According to the pricing page, the server is part of every edition: Open Source, Starter, Pro and Enterprise. The open-source edition is free, and when self-hosting you only pay for hosting. Pro and Enterprise add per-group control over the tools, MCP analytics on clients, tools and users, and browserless access through JWT, a signed sign-in token. You do not need to configure an AI provider in Metabase, because the client brings the language model. For the same reason, these calls do not count against the AI connection of Metabot, the built-in assistant.

How it differs from the alternatives

Several community servers for Metabase exist on GitHub. The built-in server stands apart mainly because it comes from the vendor, needs no external process and signs in each person through OAuth with their own permissions. There is no shared API key involved: according to the repository, the endpoint ignores API keys and only accepts OAuth tokens. Compared with a direct database MCP server, you also gain the curated models and metrics your team maintains.

Who should use the Metabase MCP Server?

It suits teams that already run Metabase and want analysis done faster, without sending every question to the data department. Data teams get a quick way to draft questions and dashboards in conversation. It is a weaker fit if the client cannot reach your instance over the network, as some clients such as Claude Desktop will not connect to localhost. Keep in mind too that query results end up in the client and, depending on your setup, may travel on to its AI provider. If personal data is involved, settle that question first.

A current Metabase instance (according to a GitHub issue, the canonical address /api/metabase-mcp applies from Metabase 62) where an admin has enabled the server under Admin > AI > MCP. AI features must not be disabled globally. The client must be able to reach the site URL. You also need a client that supports Streamable HTTP and OAuth.

With the Metabase MCP Server, your AI assistant queries data and creates questions, dashboards or alerts, always within your Metabase permissions. The server is already part of Metabase, even in the free open-source edition. This guide walks through the setup and the key safety steps.

Requirements

You need a Metabase instance and an account on it. Keep the instance up to date, because according to a GitHub issue the address /api/metabase-mcp applies from Metabase 62. An admin has to enable the server once. Your client must also be able to reach the instance. Some clients, such as Claude Desktop, will not connect to localhost.

Step 1: Enable the server

As an admin, open Admin > AI > MCP and turn on MCP server. You will find the exact address there as well. If Disable all AI features is on, the server stays off. You do not have to add an AI provider, because the language model comes from the client.

Step 2: Connect Claude Code

In the terminal, run claude mcp add --scope user --transport http metabase https://{your-metabase-url}/api/metabase-mcp. Without --scope user, the server only applies in the current directory. Then start claude, run /mcp, pick Metabase and choose Authenticate. If needed, turn the server on first with Enable. Your instance's sign-in page opens next, where you log in and confirm the permissions.

Step 3: Claude Desktop, Claude web or Cursor

In Claude, enter the same address in the connector directory and keep the defaults Sign in now and Register automatically. Leave the headers field empty. For charts inside the conversation, use chat mode in Claude Desktop and, as an admin, turn on the matching toggle under Admin > AI > MCP. In Cursor, go to Settings > MCP and add a server of type streamable-http with the same address.

First questions

It is best to begin with reading only. Try asking: Which dashboards cover revenue? What does active customer mean according to the glossary? How many orders were there last month per country? Once that works, try write requests such as creating a question or a dashboard in a separate collection.

Work safely

The assistant never has more rights than you do, but it often has more than it needs. The key risk is execute_sql, since the documentation says this tool can overwrite or delete existing data. So store read-only database accounts in Metabase wherever you can, and block the tool in the client if you do not need it. Alerts and subscriptions can send data by email or Slack, so have those actions confirmed as well. On Pro and Enterprise, admins can also allow tools per group.

claude mcp add --scope user --transport http metabase https://{your-metabase-url}/api/metabase-mcp

OAuth 2.0 through the server built into Metabase, with no external provider. The client registers itself, you sign in to your instance and confirm the scopes. According to the repository, API keys and session cookies do not work on this endpoint. Pro and Enterprise also allow browserless access through a signed JWT, and the access token then lasts one hour.

Five scopes: agent:content:read (read content and data structure), agent:content:write (create, edit and trash content), agent:query:run (run queries), agent:sql:run (run raw SQL) and agent:delivery:write (create alerts and subscriptions). Every tool stays limited to the Metabase permissions of the signed-in person.

The server runs on your Metabase instance, which means on your own infrastructure when self-hosting. Query results go to the client and, depending on your setup, on to its AI provider. For inline charts, the relevant toggles add the client's domains to the CORS allowlist. On Pro and Enterprise, MCP analytics log the individual calls.

Medium to high, depending on permissions. According to the documentation, execute_sql can overwrite or delete existing data if the database account stored in Metabase allows it. Use read-only database accounts for data sources, have agents work with restricted Metabase accounts and block execute_sql in the client. Alerts and subscriptions can send data out by email or Slack, so keep those actions behind approval too.