Postgres MCP Pro

What It Does

Connects Claude to a PostgreSQL database with tools that go far beyond running queries. Nine tools ship with the server: schema exploration (list_schemas, list_objects, get_object_details), SQL execution (execute_sql), performance analysis (explain_query, get_top_queries), index tuning (analyze_workload_indexes, analyze_query_indexes), and health monitoring (analyze_db_health).

The headline feature is index tuning. Rather than having Claude guess at what indexes might help, Postgres MCP Pro implements an adaptation of Microsoft’s Anytime Algorithm — a proven approach that explores thousands of possible index configurations using hypothetical indexes (via the hypopg extension) to find the optimal set. It shows the work at each step, including before/after query plans, so Claude can explain why each recommendation matters.

Two access modes control what Claude can do: unrestricted allows full read/write for development environments, and restricted enforces read-only transactions with execution time limits for production databases.

Built and maintained by Crystal DBA. 2K+ stars, active development, MIT licensed.

What It Does Well

• Index tuning is the killer feature. This isn’t Claude guessing “maybe add an index on created_at.” The server parses your SQL, identifies filter/join/sort columns, generates candidate indexes (including multicolumn), then uses hypopg to simulate each combination against the actual Postgres cost model. It performs cost-benefit analysis along the Pareto front — balancing performance gains against storage costs. No other Postgres MCP server does anything close to this.
• Restricted mode actually works. The server doesn’t just wrap queries in a read-only transaction and call it safe. It also parses SQL with pglast to reject statements containing COMMIT or ROLLBACK that could break out of the read-only transaction. The README is transparent about the remaining attack surface (unsafe stored procedure languages), which is refreshingly honest for a security feature.
• Health checks are deterministic, not LLM-generated. When you ask “is my database healthy?”, most approaches have Claude generate SQL to check various metrics — different every time, potentially missing things. Postgres MCP Pro runs a fixed set of validated checks adapted from PgHero: buffer cache hit rates, connection health, index health (unused/duplicate/bloated), vacuum health, sequence limits, replication lag, and constraint validation. Same checks every time, reliable results.
• Docker auto-remaps localhost. A common gotcha with database MCP servers in Docker: your database is at localhost:5432 on the host, but the container can’t reach localhost. Postgres MCP Pro automatically remaps to host.docker.internal on Mac/Windows and 172.17.0.1 on Linux. Small detail, saves real debugging time.

What It Doesn’t Do Well

• Requires PostgreSQL extensions for full value. The index tuning features need pg_stat_statements and hypopg installed on your database. On managed services (RDS, Cloud SQL, Azure) these are usually available via CREATE EXTENSION. On self-managed Postgres, pg_stat_statements needs to be in shared_preload_libraries (requires a restart), and hypopg may need to be installed from your package manager. Without these extensions, you lose the best features and are left with a competent but unremarkable SQL execution server.
• Setup difficulty is real. Between Docker, database credentials, connection URIs, access mode flags, and optional extension installation, there are more moving parts than most MCP servers. The documentation is thorough, but this isn’t a five-minute install if you’re starting from scratch. Python users can hit environment-specific issues (the README acknowledges this and recommends Docker).
• Connection is fixed at startup. You configure one database URI when the server starts, and that’s the database for the session. If you work across multiple databases, you need to reconfigure and restart. Other Postgres MCP servers like PG-MCP allow switching connections via tool calls at runtime. Both approaches have security tradeoffs — the README discusses this honestly.
• Last release was v0.3.0 in May 2025. Active repo with commits, but the release cadence is slow. The feature set is solid as-is, but if you need a fix, you may be waiting or building from source.

Setup Notes

Docker is the recommended path. The connection URI format is standard PostgreSQL: postgresql://username:password@host:port/dbname. If your database is on the same machine, use localhost — the Docker image handles the host remapping automatically.

The --access-mode flag is required. For development databases where you want Claude to create tables, insert data, etc., use --access-mode=unrestricted. For production databases where you want Claude to analyze but not modify, use --access-mode=restricted. There’s no default — you must choose explicitly, which is good design.

For the full index tuning experience, run these on your database before connecting:

CREATE EXTENSION IF NOT EXISTS pg_stat_statements;
CREATE EXTENSION IF NOT EXISTS hypopg;

The pg_stat_statements extension needs some query history to be useful. If you just installed it, let your application run for a while before asking for workload analysis.

Config

Docker with unrestricted access (development):

{
  "mcpServers": {
    "postgres": {
      "command": "docker",
      "args": [
        "run", "-i", "--rm",
        "-e", "DATABASE_URI",
        "crystaldba/postgres-mcp",
        "--access-mode=unrestricted"
      ],
      "env": {
        "DATABASE_URI": "postgresql://username:password@localhost:5432/dbname"
      }
    }
  }
}

Docker with restricted access (production):

{
  "mcpServers": {
    "postgres": {
      "command": "docker",
      "args": [
        "run", "-i", "--rm",
        "-e", "DATABASE_URI",
        "crystaldba/postgres-mcp",
        "--access-mode=restricted"
      ],
      "env": {
        "DATABASE_URI": "postgresql://username:password@localhost:5432/dbname"
      }
    }
  }
}

Python with uvx (no Docker):

{
  "mcpServers": {
    "postgres": {
      "command": "uvx",
      "args": [
        "postgres-mcp",
        "--access-mode=unrestricted"
      ],
      "env": {
        "DATABASE_URI": "postgresql://username:password@localhost:5432/dbname"
      }
    }
  }
}

Note: Don’t confuse this with the archived Anthropic reference server (@modelcontextprotocol/server-postgres). That one is read-only with no performance tools. This is postgres-mcp from crystaldba — a completely different and far more capable server.

Tested With

• Claude Desktop on Windows 11

{
  "mcpServers": {
    "postgres-mcp-pro": {
      "command": "docker",
      "args": [
        "run",
        "-i",
        "--rm",
        "-e",
        "DATABASE_URI",
        "crystaldba/postgres-mcp",
        "--access-mode=unrestricted"
      ],
      "env": {
        "DATABASE_URI": "postgresql://username:password@localhost:5432/dbname"
      }
    }
  }
}