MCP Server

turbine mcp starts a Model Context Protocol server that lets an AI agent, Claude Code, Cursor, or any MCP client, inspect your database safely. It speaks JSON-RPC 2.0 over stdio, ships with the package (no extra dependency), and exposes read-only tools only. There is no write surface and no free-form SQL input: the same safety stance as Studio.

DATABASE_URL=postgres://user:pass@localhost:5432/mydb npx turbine mcp

The server handles the MCP initialize handshake (protocol version 2025-06-18, server name turbine-orm), then serves tools/list and tools/call. Malformed input returns a JSON-RPC error and never crashes the process.

The tools#

Six read-only tools. Every database access runs inside a BEGIN READ ONLY transaction with a statement timeout, an agent physically cannot mutate your data through them.

Tool	What it returns
`schema_overview`	Tables, columns, relations, indexes, and estimated row counts for the schema.
`table_detail`	Columns, indexes, and relations for one named table.
`migrate_status`	Migration files and the tracking table, applied vs pending, without applying anything.
`doctor_report`	Missing relation indexes, from the same index advisor `turbine doctor` uses.
`explain_query`	`EXPLAIN (FORMAT JSON)` for a schema-validated `findMany` plan (see below).
`sample_rows`	Up to 50 rows from a table whose name is validated against the introspected schema. PII-tagged columns come back redacted.

explain_query never accepts free-form SQL. You pass a subset of Studio’s builder args (table + optional where / orderBy / limit / select, not with / omit). Turbine validates identifiers against the introspected schema, builds parameterized SQL via QueryInterface.buildFindMany, then runs EXPLAIN (FORMAT JSON) on that SQL with bound params only.

Argument	Required	Description
`table`	yes	Table name; must exist in the live schema.
`where`	no	`findMany`-style where clause (field names validated).
`orderBy`	no	`findMany`-style orderBy (object or array).
`limit`	no	Positive integer row limit for the planned query.
`select`	no	Field selection map (`{ field: true }`).

sample_rows caps at 50 rows, validates the table name against the live schema, and redacts PII-tagged columns before the rows reach the agent's context (the response lists which columns were redacted). Tags are a code-first declaration that introspection never infers, so the server reads them from the generated metadata in your out directory: with no generated metadata there is nothing to redact, and rows come back whole. Together, an agent can explore query plans and real data without a path to writing or to arbitrary SQL.

Options#

Option	Description
`--url`, `-u <url>`	Postgres connection string. Overrides `DATABASE_URL`.
`--schema`, `-s <name>`	Postgres schema (default: `public`).
`--include <tables>`	Comma-separated tables to include.
`--exclude <tables>`	Comma-separated tables to exclude.

Connect it to Claude Code#

Add the server to your project's .mcp.json:

{
  "mcpServers": {
    "turbine": {
      "command": "npx",
      "args": ["turbine", "mcp"],
      "env": { "DATABASE_URL": "postgres://user:pass@localhost:5432/mydb" }
    }
  }
}

Claude Code launches npx turbine mcp on startup and can then answer questions like "what indexes is the orders table missing?" or "show me a few rows from users", grounded in your real schema.

Connect it to Cursor#

Cursor uses the same shape in .cursor/mcp.json:

{
  "mcpServers": {
    "turbine": {
      "command": "npx",
      "args": ["turbine", "mcp"],
      "env": { "DATABASE_URL": "postgres://user:pass@localhost:5432/mydb" }
    }
  }
}

MCP Server

The tools#

Options#

Connect it to Claude Code#

Connect it to Cursor#

See also#