Postgres MCP

English | 🌐 Português

✨ What it does

Most LLMs interact with databases by guessing - assuming table names, inventing column names, and writing queries that may fail or expose sensitive data. Postgres MCP solves this by giving the LLM a structured, safe interface to actually understand the database before touching it.

Built with @modelcontextprotocol/sdk and pg, it provides:

• 🧠 Semantic knowledge graph - the LLM gets a complete map of schemas, tables, columns, foreign keys, inferred relations, risk levels, and business domains - built from the real schema, not invented
• 🛡️ Read-only by default - no writes, no DDL, no arbitrary SQL unless you explicitly opt in; pg_classify_query_risk lets the LLM check a query's safety before running it
• 🔐 Runtime credential resolution - credentials are read from .env at startup; nothing sensitive lives in mcp.json
• 🎯 Explicit tool selection - every tool is opt-in via tool=<name> args, so the LLM only sees what you choose to expose

📋 Requirements

• ⚙️ Node.js >= 18
• 📄 A .env file with database credentials (anywhere in the project tree - see .env Discovery)

🚀 Installation

There are two ways to use this package. Choose the one that best fits your workflow.

Option 1 - No installation (via npx, recommended for quick start)

No installation needed. npx downloads and runs the package on demand. Add -y as the first argument to skip the confirmation prompt.

{
  "servers": {
    "Postgres Tools": {
      "type": "stdio",
      "command": "npx",
      "args": [
        "-y",
        "@edelciomolina/postgres-mcp"
      ],
      "env": {
        "MCP_KEY_HOST":    "DB_HOST",
        "MCP_KEY_PORT":    "DB_PORT",
        "MCP_KEY_NAME":    "DB_NAME",
        "MCP_KEY_SSLMODE": "DB_SSLMODE",
        "MCP_KEY_USER":    "DB_USER",
        "MCP_KEY_PASS":    "DB_PASS"
      }
    }
  }
}

This starts the server with the default read-only tool set - no tool= arguments needed. To enable write-capable tools, see Write-capable tools.

💡 Using Supabase, Neon, Railway or another platform that only provides a connection string? Use MCP_KEY_URL pointing to DATABASE_URL (or whatever variable name the platform uses). The server will prioritize the URL and ignore the individual variables. See Connection via URL.

Option 2 - Install via VS Code (MCP extension marketplace)

VS Code supports discovering and installing MCP servers directly in the editor, without using the terminal.

1. Open the Command Palette (Cmd+Shift+P on Mac / Ctrl+Shift+P on Windows/Linux)
2. Run MCP: Add Server
3. Choose "Browse MCP Servers" (or "From registry", depending on your VS Code version)
4. Search for postgres-mcp or edelciomolina
5. Select Postgres MCP and follow the instructions - VS Code will add the entry to your mcp.json automatically

💡 You can also open the MCP Servers panel via Copilot chat icon → Manage MCP Servers to browse, enable, or disable servers at any time.

After installing, edit the generated entry in .vscode/mcp.json to add your tool= arguments and env key mappings as shown in the Usage section below.

🚀 Usage in VS Code (mcp.json)

Read-only (default - no tool= arguments needed):

{
  "servers": {
    "Postgres Tools": {
      "type": "stdio",
      "command": "npx",
      "args": ["@edelciomolina/postgres-mcp"],
      "env": {
        "MCP_KEY_HOST":    "DB_HOST",
        "MCP_KEY_PORT":    "DB_PORT",
        "MCP_KEY_NAME":    "DB_NAME",
        "MCP_KEY_SSLMODE": "DB_SSLMODE",
        "MCP_KEY_USER":    "DB_USER",
        "MCP_KEY_PASS":    "DB_PASS"
      }
    }
  }
}

With write tools (explicit opt-in required):

{
  "servers": {
    "Postgres Tools": {
      "type": "stdio",
      "command": "npx",
      "args": [
        "@edelciomolina/postgres-mcp",
        "tool=pg_manage_schema",
        "tool=pg_manage_indexes"
      ],
      "env": {
        "POSTGRES_MCP_ALLOW_WRITE": "true",
        "MCP_KEY_HOST":    "DB_HOST",
        "MCP_KEY_PORT":    "DB_PORT",
        "MCP_KEY_NAME":    "DB_NAME",
        "MCP_KEY_SSLMODE": "DB_SSLMODE",
        "MCP_KEY_USER":    "DB_USER",
        "MCP_KEY_PASS":    "DB_PASS"
      }
    }
  }
}

⚠️ Write-capable tools require POSTGRES_MCP_ALLOW_WRITE=true in env. Without it, the server exits at startup.

The corresponding .env file at the root of your project:

DB_HOST=db.your-project.supabase.co
DB_PORT=5432
DB_NAME=postgres
DB_SSLMODE=require
DB_USER=readonly_user
DB_PASS=your_password

⚙️ How mcp.json configuration works

🗝️ env - credential key mapping

The env block does not contain the actual credentials. It maps each MCP_KEY_* to the variable name in your .env file.

Priority: when MCP_KEY_URL (or DATABASE_URL) is present, the server uses the URL directly and ignores the individual credential keys.

This indirection lets you use any variable name in your .env - useful when sharing a .env across multiple services with different naming conventions.

🔧 args - tool selection via tool= prefix

Each enabled MCP tool is declared as a separate argument in the format tool=<name>:

"args": [
  "-y",
  "@edelciomolina/postgres-mcp",
  "tool=pg_manage_schema",
  "tool=pg_manage_indexes"
]

This makes the tool list explicit and auditable directly in mcp.json - no hidden configuration files. 🔍

🔗 Connection via URL (DATABASE_URL)

In addition to individual credentials, you can provide a full connection string - the standard format on platforms like Supabase, Neon, and Railway.

.env:

DATABASE_URL=postgresql://user:password@host:5432/database?sslmode=require

mcp.json:

{
  "servers": {
    "Postgres Tools": {
      "type": "stdio",
      "command": "npx",
      "args": ["-y", "@edelciomolina/postgres-mcp"],
      "env": {
        "MCP_KEY_URL": "DATABASE_URL"
      }
    }
  }
}

The variable mapped by MCP_KEY_URL has priority over the other keys (MCP_KEY_HOST, MCP_KEY_PORT, etc.). If the URL is present, the other variables are ignored.

If the platform uses a different name (e.g. DB_URL), just adjust the mapping:

"MCP_KEY_URL": "DB_URL"

🛡️ Why read-only is the default

If you omit all tool= arguments, the server starts with a curated read-only set - all tools that can retrieve, analyze, or explain data, but nothing that can modify it.

✅ Included in defaults (read-only):

pg_execute_query       pg_manage_query        pg_inspect_schema
pg_get_setup_instructions                     pg_analyze_database
pg_monitor_database                           pg_debug_database
pg_inspect_database_graph                     pg_describe_table_semantics
pg_find_related_tables                        pg_classify_query_risk

💡 pg_execute_query rejects INSERT, UPDATE, DELETE, DDL, ANALYZE, VACUUM, EXPLAIN ANALYZE and other write/maintenance commands before the database is queried.

💡 pg_inspect_schema provides read-only schema introspection (get_info, get_enums). For DDL operations, use pg_manage_schema with explicit opt-in.

⚠️ Excluded from defaults - require tool= argument AND POSTGRES_MCP_ALLOW_WRITE=true:

📍 .env file discovery

The server resolves the .env file in this order:

1. env-file=<path> argument - explicit path relative to cwd; takes priority over everything
2. Upward search - starting from cwd, searches each parent directory until a .env is found or the filesystem root is reached

If no .env is found, the server exits with a clear error message.

Monorepos and subfolders

When VS Code starts the MCP process, cwd is typically the workspace root. If your .env is in a subfolder (e.g. functions/.env), use env-file= to point to it explicitly:

{
  "servers": {
    "Postgres Tools": {
      "type": "stdio",
      "command": "npx",
      "args": [
        "-y",
        "@edelciomolina/postgres-mcp",
        "env-file=functions/.env"
      ],
      "env": {
        "MCP_KEY_HOST":    "DB_HOST",
        "MCP_KEY_PORT":    "DB_PORT",
        "MCP_KEY_NAME":    "DB_NAME",
        "MCP_KEY_SSLMODE": "DB_SSLMODE",
        "MCP_KEY_USER":    "DB_USER",
        "MCP_KEY_PASS":    "DB_PASS"
      }
    }
  }
}

💡 The upward search behavior handles the common case automatically. Use env-file= when you need explicit control (CI, monorepos, Docker bind-mounts).

🧰 Available tools

Read-only (enabled by default)

Write-capable (opt-in via tool= argument + POSTGRES_MCP_ALLOW_WRITE=true)

🧠 Semantic Layer

The four pg_*_graph / pg_*_semantics / pg_*_risk tools build an in-memory knowledge graph of your database at runtime. This gives the LLM a structured map - schemas, tables, columns, foreign keys, inferred relations, risk levels, and business domains - without executing any query against your data.

All inferred fields (column semantic roles, table probable types, inferred relations) are clearly tagged so the LLM knows to treat them as hints, not schema facts.

Optional configuration (mcp-config.json)

Place a mcp-config.json file beside your .env to tune the semantic layer and security limits. All fields are optional - omitting the file applies safe defaults.

{
  "security": {
    "defaultLimit": 100,
    "maxLimit": 1000,
    "blockedSchemas": ["pg_catalog", "information_schema"],
    "blockedTables": [],
    "requireLimit": true
  },
  "semanticLayer": {
    "enabled": true,
    "inferRelationsWithoutForeignKeys": true,
    "inferBusinessEntities": true,
    "sensitiveKeywords": ["password", "secret", "token", "api_key", "ssn", "hash"]
  }
}

🏗️ Architecture

For a detailed view of the communication flow between the MCP client, the proxy, and PostgreSQL - including the full sequence diagram - see ARCHITECT.md.

📄 License

MIT © Edelcio Molina