Skip to main content
Glama
db-mcp
by eklemen

db-mcp — Read-Only Database MCP Server

A Model Context Protocol server that lets AI assistants safely explore a PostgreSQL database — schema, tables, columns, relationships, indexes, constraints, DDL, and sample data — without ever modifying it.

Safety model

Read-only is enforced at three independent layers, so a failure in any one layer is still caught by the others:

  1. Least-privilege database role. You connect with a role that only has CONNECT / USAGE / SELECT. The database itself rejects anything else.

  2. Read-only transaction. Every query runs inside BEGIN TRANSACTION READ ONLY on a session forced to default_transaction_read_only = on with a statement_timeout. The query is always rolled back.

  3. Application-level validation. Arbitrary SQL (the run_readonly_query tool) is parsed with libpg-query — the real PostgreSQL parser compiled to wasm — and rejected unless it is a single read-only statement. This catches multi-statement input, DDL/DML, EXPLAIN ANALYZE, SELECT INTO, and data-modifying CTEs (WITH x AS (INSERT ... RETURNING ...) SELECT ...).

Additional guards: a maximum query length, a hard row cap on every result, a sample-row cap (default 50), and secrets are never logged (config is redacted before any diagnostic output).

Related MCP server: postgres-mcp-readonly

Tools

Tool

Description

list_databases

Databases the read-only role can connect to.

list_schemas

Schemas visible to the role (system schemas excluded).

list_tables

Tables/views in a schema: type, row estimate, comment.

describe_table

Columns, PK, FKs, unique/check constraints, indexes.

get_table_ddl

Reconstructed CREATE TABLE DDL (+ indexes).

get_relationships

Foreign keys, referenced tables, cardinality hint.

get_indexes

Index names, columns, order, uniqueness, type, predicate.

sample_rows

A small sample of rows (capped at 50).

run_readonly_query

Validated read-only SQL; returns columns, rows, timing.

search_schema

Search names + comments across tables/columns/constraints/indexes.

Requirements

  • Node.js >= 18.17

  • A reachable PostgreSQL database and a read-only role (see below).

Setup

npm install
npm run build

Create a least-privilege read-only role

Run as a superuser or the database owner:

CREATE ROLE readonly_user LOGIN PASSWORD 'CHANGE_ME';
GRANT CONNECT ON DATABASE your_db TO readonly_user;
GRANT USAGE ON SCHEMA public TO readonly_user;
GRANT SELECT ON ALL TABLES IN SCHEMA public TO readonly_user;
ALTER DEFAULT PRIVILEGES IN SCHEMA public GRANT SELECT ON TABLES TO readonly_user;
-- Optional extra safety:
ALTER ROLE readonly_user SET default_transaction_read_only = on;

Configure

Copy .env.example and fill in your connection. Configuration is read from the environment (no .env is loaded automatically — export the vars, or have your MCP client pass them in its env block).

Variable

Default

Purpose

DATABASE_URL

—

Full connection string (wins over PG*).

PGHOST / PGPORT / PGDATABASE / PGUSER / PGPASSWORD

—

Discrete connection settings.

PGSSLMODE

disable

require/verify-* enable TLS.

DB_MCP_DRIVER

postgres

Database driver.

DB_MCP_MAX_ROWS

1000

Hard cap for run_readonly_query.

DB_MCP_SAMPLE_MAX_ROWS

50

Hard cap for sample_rows.

DB_MCP_STATEMENT_TIMEOUT_MS

5000

Per-statement timeout.

DB_MCP_MAX_QUERY_LENGTH

10000

Reject longer queries.

DB_MCP_DEFAULT_SCHEMA

public

Schema used when omitted.

Running locally

export DATABASE_URL="postgres://readonly_user:***@localhost:5432/your_db"
npm start

The server speaks MCP over stdio. To poke at it interactively, use the MCP Inspector:

npx @modelcontextprotocol/inspector node dist/server.js

Connecting from Claude Code

Add an entry to your MCP config (e.g. .mcp.json in your project, or the global Claude Code config):

{
  "mcpServers": {
    "db-mcp": {
      "command": "node",
      "args": ["/absolute/path/to/db-mcp/dist/server.js"],
      "env": {
        "DATABASE_URL": "postgres://readonly_user:***@localhost:5432/your_db"
      }
    }
  }
}

Then ask Claude to, e.g., "list the tables in the public schema" or "show me the DDL for the orders table".

Tests

npm test

Tests are pure unit tests (no live database required): exhaustive SQL-validation cases (allowed reads vs. rejected mutations and bypass attempts), identifier quoting, DDL reconstruction, and config parsing/redaction.

Extending to other databases

Tools depend only on the DatabaseDriver interface (src/db/driver.ts). To add MySQL, SQLite, or Snowflake:

  1. Implement DatabaseDriver for the new engine under src/db/<engine>/.

  2. Register it in createDriver (src/db/factory.ts).

  3. Provide an engine-appropriate read-only enforcement (read-only role + read-only session/transaction) and SQL validation.

Project layout

src/
  server.ts                 MCP server setup + stdio transport
  config.ts                 env parsing + secret redaction
  errors.ts                 typed error classes
  types.ts                  shared result types
  validation/
    validateReadOnly.ts     libpg-query-based read-only enforcement
  db/
    driver.ts               DatabaseDriver interface
    factory.ts              driver selection
    postgres/
      pool.ts               read-only pool, transactions, row-capped cursor
      identifiers.ts        safe identifier quoting
      introspection.ts      catalog/information_schema queries
      ddl.ts                CREATE TABLE reconstruction (pure)
      PostgresDriver.ts     DatabaseDriver implementation
  tools/                    one module per MCP tool + registry
tests/                      pure unit tests

Available Tools

10 tools
describe_tableA

Describe a table: columns (type, nullability, default, comment), primary key, foreign keys, unique constraints, check constraints, and indexes.

Inputs: table (required), schema.

get_indexesA

List indexes on a table: name, access method (btree/gin/gist/...), uniqueness, whether it backs the primary key, ordered columns with sort direction and NULLS ordering, and any partial-index predicate.

Inputs: table (required), schema.

get_relationshipsA

List foreign-key relationships with constraint names, referenced tables/columns, ON UPDATE/DELETE actions, and a cardinality hint (1:1 when the FK columns are uniquely constrained locally, otherwise many:1). Optionally filter by schema and/or table.

Inputs: table, schema.

get_table_ddlA

Return the reconstructed CREATE TABLE DDL for a table (PostgreSQL has no SHOW CREATE TABLE, so this is the closest supported reconstruction from the catalog, including constraints and indexes).

Inputs: table (required), schema.

list_databasesA

List databases the read-only user can connect to, with owner and comment when available.

Inputs: no named parameters.

list_schemasA

List schemas (namespaces) the read-only user can access. Excludes system schemas (pg_*, information_schema).

Inputs: no named parameters.

list_tablesA

List tables and views in a schema. Returns name, type (table/view/materialized view/foreign table), a planner row estimate when available, and any comment. Defaults to the configured default schema.

Inputs: schema.

run_readonly_queryA

Execute a single read-only SQL query (SELECT / EXPLAIN-without-ANALYZE / SHOW). The query is validated with a real PostgreSQL parser and rejected if it contains any write, DDL, data-modifying CTE, multiple statements, or other side effect. Results are capped at the server row maximum. Returns columns, rows, row count, execution time, and a truncation flag.

Inputs: sql (required), limit.

sample_rowsA

Return a small sample of rows from a table. The limit is capped at the server's hard maximum (default 50) regardless of the requested value.

Inputs: table (required), limit, schema.

search_schemaA

Search across the schema for a term (case-insensitive substring). Matches table names, column names, table/column comments, constraint names, and index names. Optionally restrict to a single schema.

Inputs: term (required), schema.

Tool Schema Changelog

Recent tool additions, removals, and schema changes observed during successful MCP inspections.

Loading tool schema changelog…

    TDQS

    A4.1/5.0

    Scored across 10 tools

    Disambiguation5/5

    Each tool has a clearly distinct purpose: describing table structure, fetching indexes, listing foreign keys, retrieving DDL, enumerating databases/schemas/tables, running queries, sampling rows, and searching schema metadata. No two tools overlap in functionality.

    Naming Consistency5/5

    All tools follow a consistent verb_noun pattern using snake_case (e.g., describe_table, list_schemas, run_readonly_query). The verbs accurately describe the action (describe, get, list, run, sample, search), and there are no mixed naming conventions.

    Tool Count5/5

    With 10 tools, the server is well-scoped for database introspection and read-only queries. Each tool addresses a distinct need without redundancy or excessive granularity, fitting the typical range of 3-15 tools for a focused domain.

    Completeness4/5

    The tool set covers core CRUD-like operations for database metadata (create via DDL retrieval, read via descriptions and queries, and listing). Minor gaps exist, such as missing explicit tools for table size estimation or listing extensions, but most common exploration workflows are supported.

    Related MCP Connectors

    Related MCP Servers

    • A
      license
      Not graded
      quality
      D
      maintenance
      Read-only PostgreSQL MCP server that enables running SELECT queries, listing tables and schemas, and describing columns, with built-in protection against writes and malicious SQL attacks.
      432 npm
      MIT
    • A
      license
      Not graded
      quality
      D
      maintenance
      Read-only MCP server for PostgreSQL, enabling schema discovery, table metadata, and safe SELECT queries via READ ONLY transactions.
      12 npm
      MIT
    • A
      license
      Not graded
      quality
      D
      maintenance
      Provides a read-only PostgreSQL MCP server with schema introspection. Enforces least-privilege database roles to prevent any writes, even from malicious SQL.
      MIT