PostgreSQL README

MCP PostgreSQL Server

A Model Context Protocol (MCP) server for PostgreSQL: local, Docker, RDS, Neon, and Supabase databases.

The server is small and auditable, with four runtime dependencies: the MCP SDK, pg, pg-connection-string, and zod (plus ssh2, an optional dependency used only for SSH tunneling).

Requires Node.js 20 or newer.

Quick start

The preferred way to configure the server is a single DATABASE_URL:

{
  "mcpServers": {
    "postgres": {
      "type": "stdio",
      "command": "npx",
      "args": ["-y", "mcp-postgres-server"],
      "env": {
        "DATABASE_URL": "postgres://user:password@localhost:5432/mydb",
        "PG_ALLOW_WRITE": "false"
      }
    }
  }
}

With PG_ALLOW_WRITE set to "false" the server has read-only access to the database. This is the default; set it to "true" only if the model must write.

The same JSON works in any MCP client that speaks stdio: VS Code, Cursor, Claude Code, Codex, Windsurf.

Alternatively, set the individual PG_* variables; they are used when DATABASE_URL is not set:

{
  "mcpServers": {
    "postgres": {
      "type": "stdio",
      "command": "npx",
      "args": ["-y", "mcp-postgres-server"],
      "env": {
        "PG_HOST": "your_host",
        "PG_PORT": "5432",
        "PG_USER": "your_user",
        "PG_PASSWORD": "your_password",
        "PG_DATABASE": "your_database",
        "PG_ALLOW_WRITE": "false"
      }
    }
  }
}

Manual Installation

npm install mcp-postgres-server

Or run directly with:

npx mcp-postgres-server

Connect to your database

Local Postgres:

DATABASE_URL=postgres://mcp_readonly:secret@localhost:5432/mydb

Postgres in Docker: if the database runs in a container with a published port, connect to localhost:<published-port> as usual. If the MCP server itself runs inside a container and the database runs on your host machine, use host.docker.internal instead of localhost:

DATABASE_URL=postgres://mcp_readonly:secret@host.docker.internal:5432/mydb

Amazon RDS:

DATABASE_URL=postgres://mcp_readonly:secret@mydb.xxxxxx.us-east-1.rds.amazonaws.com:5432/mydb?sslmode=require

Neon:

DATABASE_URL=postgres://mcp_readonly:secret@ep-xxx-xxx.us-east-2.aws.neon.tech/mydb?sslmode=require

Supabase:

DATABASE_URL=postgres://postgres.xxxxxxxx:secret@aws-0-us-east-1.pooler.supabase.com:5432/postgres?sslmode=require

Tools

Tool availability depends on configuration:

1. query

Execute a read-only SQL statement. Accepts SELECT, WITH ... SELECT, EXPLAIN, and SHOW. One statement per call - multi-statement input is rejected by the extended query protocol. In read-only mode (the default) the statement runs as BEGIN READ ONLY, the query, and ROLLBACK - three commands, roughly two network round trips with pipelining - so the database itself refuses any write. With PG_ALLOW_WRITE=true the statement is sent directly, without that wrapper, so a write run through query would execute - use execute for writes. Supports PostgreSQL-style $1, $2 prepared-statement parameters; values are bound by the driver and never spliced into the SQL text.

use_mcp_tool({
  server_name: "postgres",
  tool_name: "query",
  arguments: {
    sql: "SELECT * FROM users WHERE id = $1",
    params: [1]
  }
});

Returns compact JSON: {"rows": [...], "rowCount": n, "returnedRows": n, "truncated": false}. When the serialized rows exceed PG_MAX_RESULT_BYTES, only the rows that fit are returned (returnedRows < rowCount), truncated is true, and a hint suggests adding LIMIT/WHERE or selecting fewer columns.

2. list_schemas

List all schemas in the connected database.

use_mcp_tool({
  server_name: "postgres",
  tool_name: "list_schemas",
  arguments: {}
});

3. list_tables

List tables in the connected database. Accepts an optional schema parameter (defaults to 'public').

// List tables in the 'public' schema (default)
use_mcp_tool({
  server_name: "postgres",
  tool_name: "list_tables",
  arguments: {}
});

// List tables in a specific schema
use_mcp_tool({
  server_name: "postgres",
  tool_name: "list_tables",
  arguments: {
    schema: "my_schema"
  }
});

4. describe_table

Get the structure of a specific table (columns, types, nullability, defaults, primary keys). Accepts an optional schema parameter (defaults to 'public').

use_mcp_tool({
  server_name: "postgres",
  tool_name: "describe_table",
  arguments: {
    table: "users",
    schema: "my_schema"  // optional
  }
});

5. execute - requires PG_ALLOW_WRITE=true

Execute an INSERT, UPDATE, DELETE, or DDL statement. Always registered, but in read-only mode (the default) it refuses with an error naming PG_ALLOW_WRITE and changes nothing - the statement never reaches the database. With PG_ALLOW_WRITE=true it runs: same $1, $2 parameter handling as query, one complete statement per call, and the connecting role governs what it may do. Returns {"rowCount": n, "command": "INSERT"}.

use_mcp_tool({
  server_name: "postgres",
  tool_name: "execute",
  arguments: {
    sql: "INSERT INTO users (name, email) VALUES ($1, $2)",
    params: ["John Doe", "john@example.com"]
  }
});

6. connect_db - requires PG_ENABLE_RUNTIME_CONNECT=true

Connect to a different PostgreSQL database at runtime using provided credentials. Not registered by default - prefer configuring credentials through the environment so they never pass through model-visible arguments. Session limits (statement_timeout, idle_in_transaction_session_timeout) are re-applied after every reconnect; read-only reads enforce read-only in their own BEGIN READ ONLY transaction.

use_mcp_tool({
  server_name: "postgres",
  tool_name: "connect_db",
  arguments: {
    host: "localhost",
    port: 5432,
    user: "your_user",
    password: "your_password",
    database: "your_database"
  }
});

Configuration reference

To reach a database only accessible through a bastion, see SSH tunneling (adds PG_SSH_* variables).

Features

• Read-only by default; writes are an explicit opt-in (PG_ALLOW_WRITE=true)
• Read-only enforced by the engine (BEGIN READ ONLY), never by client-side SQL parsing
• Data access behind a small typed interface; the pg driver never leaks past it
• DATABASE_URL support with SSL (sslmode=disable|allow|prefer|require|verify-ca|verify-full, custom CA)
• Prepared-statement parameters: $1-style placeholders, bound by the driver
• Result size cap (byte budget) with an explicit truncated flag instead of flooding the model's context
• Session statement timeout plus a client deadline; transaction poolers may not preserve session settings
• Errors returned as readable tool results with SQLSTATE-based hints, so the model can self-correct
• Survives dropped connections - reconnects lazily instead of crashing
• Optional SSH tunneling (PG_SSH_*) with mandatory host-key verification, loaded only when configured
• MCP tool annotations (read-only / destructive hints) per spec 2025-11-25
• Multi-schema support for database operations

Security

Full details, including the threat model and disclosure process, are in SECURITY.md. The short version:

1. A least-privilege database role is the real boundary. The MCP works with existing credentials; creating or changing roles is not required. A dedicated role is what actually guarantees writes are impossible. On PostgreSQL 14+, the following is a starting point:

   sqlCopy

   CREATE ROLE mcp_readonly LOGIN PASSWORD 'change-me';
   GRANT CONNECT ON DATABASE your_database TO mcp_readonly;
   GRANT pg_read_all_data TO mcp_readonly;                          -- adds read privileges
   ALTER ROLE mcp_readonly SET default_transaction_read_only = on;  -- read-only by default
   

   (On PostgreSQL 13 or older, grant SELECT explicitly instead of pg_read_all_data - see SECURITY.md.) The server warns on stderr if you connect as a superuser. Read grants do not revoke existing privileges, and defaults remain mutable; available functions, ownership and inherited privileges also matter.

2. The engine enforces read-only. There is no client-side SQL parsing. In read-only mode every read runs in a rolled-back BEGIN READ ONLY transaction, so PostgreSQL itself - which alone knows what a function, view, or rule does - refuses any write with SQLSTATE 25006 and reverts any session change the statement made. The extended protocol rejects multi-command strings.

Honest framing: the read-only transaction is defense-in-depth on top of the role, not a replacement for it. Read-only mode stops a confused or prompt-injected model from writing to your database; it does not stop prompt injection carried in the row data a query returns. Don't point this server at production - use a replica, a snapshot, or a tightly scoped role. See SECURITY.md.

SSH tunneling

Set PG_SSH_HOST (plus auth and host-key verification) to reach a database that is only accessible through a bastion (an SSH jump host). The connection string / PG_* fields then describe the database as seen from the bastion:

{
  "mcpServers": {
    "postgres": {
      "type": "stdio",
      "command": "npx",
      "args": ["-y", "mcp-postgres-server"],
      "env": {
        "DATABASE_URL": "postgres://mcp_readonly:secret@db.internal:5432/mydb?sslmode=verify-full",
        "PG_SSH_HOST": "bastion.example.com",
        "PG_SSH_USER": "jump",
        "PG_SSH_PRIVATE_KEY": "/home/me/.ssh/id_ed25519",
        "PG_SSH_FINGERPRINT": "SHA256:xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx"
      }
    }
  }
}

• SSH changes only the transport. Read-only enforcement, the result size cap, timeouts, and connect_db behave exactly as on a direct connection, and no extra SQL is sent per query.
• Host-key verification is mandatory via a pinned PG_SSH_FINGERPRINT - the tunnel will not connect without it, so a man-in-the-middle bastion is refused. Get the fingerprint over a channel you trust, most trustworthy first:
  • on the bastion itself, or from its admin: ssh-keygen -lf /etc/ssh/ssh_host_ed25519_key.pub (no network involved);
  • from your existing ~/.ssh/known_hosts, if you already reach the host over ssh: ssh-keygen -lF bastion.example.com;
  • fetched from the host: ssh-keyscan bastion.example.com | ssh-keygen -lf - (trust this only when run from a network position you trust - it accepts whatever the host returns).
• TLS validates the real database hostname. With verify-full, the certificate is checked against the database's own hostname (e.g. db.internal), not the loopback the tunnel binds locally, and rejectUnauthorized is pinned on so an inherited NODE_TLS_REJECT_UNAUTHORIZED=0 cannot disable it.
• ssh2 is an optional dependency, loaded only when PG_SSH_HOST is set, so a direct connection never initializes it. npm installs optional dependencies by default; run npm install --omit=optional to skip it entirely (a direct connection does not need it).

A tunneled connection that fails reports a stable SSH_* code - see Error Handling.

Error Handling

SQL and connection failures are returned as tool results (isError: true) with a message, the SQLSTATE code, and a hint. PostgreSQL's own server hint is used when present; otherwise these fallbacks apply:

Over an SSH tunnel, a failure carries a stable code (and, where the cause is determinate, a hint naming the setting to fix), so the failing phase is unambiguous:

A genuine PostgreSQL error through a healthy tunnel keeps its own code (e.g. 28P01 for wrong database credentials), not an SSH code.

Migrating from 0.1.x

Not needed for new installs. Two behavior changes since 0.1.x:

1. Read-only by default. The execute tool is always visible but refuses writes (with an error naming the flag) unless PG_ALLOW_WRITE=true, and every read runs inside an engine-enforced READ ONLY transaction. If your workflow writes to the database, set "PG_ALLOW_WRITE": "true" to restore 0.1.x behavior.
2. connect_db is disabled by default. Runtime connection switching (passing credentials through tool arguments) requires PG_ENABLE_RUNTIME_CONNECT=true; otherwise connection details come only from the environment.

Tool names, parameter names, and PG_* variables are unchanged. Result payloads are now structured compact JSON for every tool (e.g. query returns {rows, rowCount, returnedRows, truncated} instead of a bare row array) - see CHANGELOG.md for the exact shapes before updating anything that parses tool output.

License

MIT