Starburst MCP Server

Starburst's MCP server exposes the Starburst analytics platform (built on Trino) to AI agents and LLM applications. It is available in two flavors: a fully managed, multi-tenant service hosted inside Starburst Galaxy's control plane, and an integrated MCP server that ships with Starburst Enterprise for self-hosted clusters. In both cases, agents connect over HTTP and submit SQL strings, receiving structured JSON results with column metadata and rows.

The server is intentionally read-only. SQL parsing rejects any data-modifying statements (INSERT, UPDATE, DELETE, MERGE, TRUNCATE, CREATE, ALTER, DROP, GRANT, REVOKE) before queries reach the engine, leaving only SELECT, SHOW, and non-destructive EXPLAIN operations. All requests inherit Starburst's existing role-based and attribute-based access controls through Starburst Gravity, and every query is audit-logged with origin metadata (source tag, client user-agent, trace token).

What makes it notable is the federated reach: a single MCP endpoint lets an agent query across any source Trino connects to (S3/Iceberg, Snowflake, Postgres, MySQL, Kafka, Delta Lake, and more) without copying data. The Enterprise build adds higher-level tools beyond raw SQL, including data product search, parameterized query templates, and a stateful AI agent tool scoped to a data product.

Tools

Prerequisites

• A Starburst Galaxy account (Mission Critical, Enterprise, or free trial) OR a Starburst Enterprise cluster with a valid MCP license
• An MCP-compatible client (Claude Desktop, Cursor, VS Code, etc.)
• Credentials: either Galaxy username/password, a service account, or an OAuth client with the galaxy.mcp scope

Starburst Galaxy (hosted) endpoint

The hosted endpoint follows this pattern:

https://<account-name>.mcp.galaxy.starburst.io

Claude Desktop config (Basic Auth)

Base64-encode email:password and add it to your Claude Desktop config:

{
  "mcpServers": {
    "starburst-galaxy": {
      "type": "http",
      "url": "https://<account-name>.mcp.galaxy.starburst.io",
      "headers": {
        "Authorization": "Basic <base64_encoded_credentials>"
      }
    }
  }
}

Starburst Enterprise (self-hosted)

Enable the MCP server on the coordinator:

starburst.mcp.enabled=true

Then point clients at:

https://<coordinator-host>:<port>/mcp

Authentication uses standard Starburst auth (OAuth 2.0, delegated OAuth, or HTTP headers per the Client Protocol spec).

Notes and limits

• Galaxy result sets are capped at 100KB; Enterprise defaults to 1MB (configurable 64KB to 8GB)
• Enterprise query execution defaults to a 5 minute max
• OAuth Dynamic Client Registration is not supported on Galaxy
• Custom roles cannot be set for MCP operations; users see only resources their role permits

• Let an analyst agent answer ad hoc business questions by writing SELECT queries across federated sources (S3/Iceberg + Snowflake + Postgres) without moving data
• Have an agent discover the right governed dataset via searchDataProducts, inspect its schema with getDataProductDetails, then run a query against it
• Expose curated, parameterized queries (e.g., "revenue by region for date X") that agents can invoke safely without writing raw SQL
• Build a multi-agent analytics workflow where each agent has scoped, audit-logged access to specific data products under Starburst Gravity RBAC/ABAC
• Use Claude or another LLM client to run SHOW CATALOGS, SHOW SCHEMAS, and EXPLAIN to explore a Trino-backed data mesh and document its structure

• "Show me total orders by region for the last 30 days using the sales catalog in Starburst."
• "Search Starburst data products for anything related to customer churn and summarize the top three matches."
• "List the catalogs and schemas in my Galaxy cluster, then describe the columns in the analytics.events table."
• "Run the monthly_revenue_by_segment parameterized query for segment = 'enterprise' and month = '2026-04'."
• "Explain the query plan for SELECT customer_id, SUM(amount) FROM iceberg.sales.orders GROUP BY 1 and tell me which join strategy Trino will use."

• Official, vendor-hosted MCP service for Galaxy customers, with no local binary to install or maintain
• Federated reach across every source Trino can connect to, queried in place without data copies
• Enforced read-only semantics plus full RBAC/ABAC, audit logging, and trace tagging through Starburst Gravity
• Enterprise build adds higher-level tools (data product search, parameterized queries, AI agent sessions) beyond raw SQL

• No write operations of any kind: INSERT, UPDATE, DELETE, and DDL are all rejected
• Galaxy result sets capped at 100KB, which can truncate large analytical responses
• Galaxy MCP requires a paid Mission Critical/Enterprise plan (or trial); Enterprise MCP requires a licensed cluster and explicit enablement
• OAuth Dynamic Client Registration and custom MCP-scoped roles are not supported on Galaxy

• Trino MCP server (community implementations) for self-hosted open-source Trino clusters
• Snowflake's official MCP/Cortex integrations for Snowflake-native analytics
• Databricks Genie / SQL MCP integrations for Lakehouse-based federated SQL