Metabase MCP

What 1luvc0d3/metabase-mcp MCP server does

The 1luvc0d3/metabase-mcp MCP server connects an MCP client to a Metabase instance using a Metabase API key. Its tool surface covers reading dashboards, cards, and databases; creating and updating questions, dashboards, and collections; executing SELECT-oriented SQL; and deleting cards or dashboards. It also supports adding and removing cards from dashboards.

The package includes 30 tools across read, batch, workflow, write, natural-language query, and insight functions. Optional Anthropic integration enables natural-language questions that produce SQL and results, along with SQL explanation, optimization, validation, automated insights, and trend analysis. These AI features require an Anthropic API key.

How it works

The server runs locally as an MCP process and sends requests to the configured Metabase URL. API-key authentication makes it suitable for CI and autonomous agents without an OAuth browser flow. It can also connect to self-hosted Metabase deployments that are reachable from the host running the process, including instances behind a VPN.

Three modes control the broad permission set: read is the default, write enables write operations, and full includes AI insight capabilities. MCP_TOOLS_ALLOW and MCP_TOOLS_DENY provide per-tool filtering; deny rules take precedence and also apply to tools nested inside batches and workflows.

batch_execute can run up to 20 parallel operations. run_workflow chains steps and lets later steps reference earlier outputs with expressions such as $stepName.path. Responses use compact JSON by default, while format: "default" requests expanded output. Schema caching supplies context for natural-language queries against larger databases.

Setup and configuration

Install the npm package with npx @ai-1luvc0d3/metabase-mcp, install it globally with npm, or build and start the project from source. Claude Desktop can also install a release .mcpb package and prompt for the Metabase URL and API key, storing those values in the operating system keychain.

The required variables are METABASE_URL and METABASE_API_KEY. Set MCP_MODE to read, write, or full as needed. ANTHROPIC_API_KEY is optional and enables the server-side AI functions. Other settings control request timeout, maximum returned rows, log level, tool allow/deny lists, and separate read, write, and LLM request limits.

Tools and capabilities

SQL execution includes injection detection, DDL/DML blocking, and dangerous-pattern enforcement. Rate limits are configurable independently for read, write, and LLM operations. Each operation is audit-logged with a risk assessment. These controls make the 1luvc0d3/metabase-mcp MCP server suited to deployments where agents need access to BI data without unrestricted tool exposure.

Limitations and notes

Authentication uses an API key rather than OAuth, so the server does not provide OAuth-based per-user permission scoping. It does not provide inline interactive charts rendered in the AI client. The optional natural-language and insight functions depend on Anthropic, and raw query access remains subject to the server's SQL guardrails and configured row limit.

The project is intended to work with Metabase versions older than 0.60 as well as newer versions. Choose the official Metabase MCP server instead when OAuth-based user scoping or inline interactive charts are more important than headless API-key access and the controls provided by the 1luvc0d3/metabase-mcp MCP server.