Pgtuner MCP

🚀 PostgreSQL Performance Tuning MCP

A Model Context Protocol (MCP) server that offers AI-driven PostgreSQL performance tuning. It helps identify slow queries, recommend optimal indexes, analyze execution plans, and use HypoPG for hypothetical index testing.

🚀 Quick Start

This MCP server is designed to provide AI-powered PostgreSQL performance tuning. To get started, you can follow the installation and configuration steps below.

✨ Features

Query Analysis

• Retrieve slow queries from pg_stat_statements along with detailed statistics.
• Analyze query execution plans using EXPLAIN and EXPLAIN ANALYZE.
• Automatically identify performance bottlenecks through plan analysis.
• Monitor active queries and detect long - running transactions.

Index Tuning

• Provide AI - powered index recommendations based on query workload analysis.
• Conduct hypothetical index testing with the HypoPG extension (without disk usage).
• Find unused and duplicate indexes for cleanup.
• Estimate index sizes before creation.
• Test query plans with proposed indexes before implementation.

Database Health

• Offer comprehensive health scoring through multiple checks.
• Monitor connection utilization.
• Analyze cache hit ratios (buffer and index).
• Detect lock contention.
• Monitor vacuum health and transaction ID wraparound.
• Monitor replication lag.
• Analyze background writer and checkpoint I/O.

Vacuum Monitoring

• Track long - running VACUUM and VACUUM FULL operations in real - time.
• Monitor autovacuum progress and performance.
• Identify tables that need vacuuming.
• View recent vacuum activity history.
• Analyze the effectiveness of autovacuum configuration.

I/O Performance Analysis

• Analyze disk read/write patterns across tables and indexes.
• Identify I/O bottlenecks and hot tables.
• Monitor buffer cache hit ratios.
• Track temporary file usage indicating work_mem issues.
• Analyze checkpoint and background writer I/O.
• Support enhanced pg_stat_io metrics for PostgreSQL 16+.

Configuration Analysis

• Review PostgreSQL settings by category.
• Provide recommendations for memory, checkpoint, WAL, autovacuum, and connection settings.
• Identify suboptimal configurations.

MCP Prompts & Resources

• Offer pre - defined prompt templates for common tuning workflows.
• Provide dynamic resources for table stats, index info, and health checks.
• Supply comprehensive documentation resources.

📦 Installation

Standard Installation (for MCP clients like Claude Desktop)

pip install pgtuner_mcp

Or using uv:

uv pip install pgtuner_mcp

Manual Installation

git clone https://github.com/isdaniel/pgtuner_mcp.git
cd pgtuner_mcp
pip install -e .

📚 Documentation

Configuration

Environment Variables

Connection String Format: postgresql://user:password@host:port/database

Minimal User Permissions

To run this MCP server, the PostgreSQL user needs specific permissions to query system catalogs and extensions. Here are the minimal permissions required for different feature sets.

Basic Permissions (Required for Core Functionality)

-- Create a dedicated monitoring user
CREATE USER pgtuner_monitor WITH PASSWORD 'secure_password';

-- Grant connection to the target database
GRANT CONNECT ON DATABASE your_database TO pgtuner_monitor;

-- Grant usage on schemas
GRANT USAGE ON SCHEMA public TO pgtuner_monitor;
GRANT USAGE ON SCHEMA pg_catalog TO pgtuner_monitor;

-- Grant SELECT on user tables and indexes (for table stats and analysis)
GRANT SELECT ON ALL TABLES IN SCHEMA public TO pgtuner_monitor;
ALTER DEFAULT PRIVILEGES IN SCHEMA public GRANT SELECT ON TABLES TO pgtuner_monitor;

-- Grant access to system catalog views (read-only)
GRANT pg_read_all_stats TO pgtuner_monitor;  -- PostgreSQL 10+

Extension - Specific Permissions

For pgstattuple (Bloat Detection):

-- Create the extension (requires superuser or appropriate privileges)
CREATE EXTENSION IF NOT EXISTS pgstattuple;

-- Grant execution on pgstattuple functions
GRANT EXECUTE ON FUNCTION pgstattuple(regclass) TO pgtuner_monitor;
GRANT EXECUTE ON FUNCTION pgstattuple_approx(regclass) TO pgtuner_monitor;
GRANT EXECUTE ON FUNCTION pgstatindex(regclass) TO pgtuner_monitor;
GRANT EXECUTE ON FUNCTION pgstatginindex(regclass) TO pgtuner_monitor;
GRANT EXECUTE ON FUNCTION pgstathashindex(regclass) TO pgtuner_monitor;

-- Alternative: Use pg_stat_scan_tables role (PostgreSQL 14+)
GRANT pg_stat_scan_tables TO pgtuner_monitor;

For HypoPG (Hypothetical Index Testing):

-- Create the extension (requires superuser or appropriate privileges)
CREATE EXTENSION IF NOT EXISTS hypopg;

-- Grant SELECT on HypoPG views
GRANT SELECT ON hypopg_list_indexes TO pgtuner_monitor;
GRANT SELECT ON hypopg_hidden_indexes TO pgtuner_monitor;

-- Grant execution on HypoPG functions with proper signatures
GRANT EXECUTE ON FUNCTION hypopg_create_index(text) TO pgtuner_monitor;
GRANT EXECUTE ON FUNCTION hypopg_drop_index(oid) TO pgtuner_monitor;
GRANT EXECUTE ON FUNCTION hypopg_reset() TO pgtuner_monitor;
GRANT EXECUTE ON FUNCTION hypopg_hide_index(oid) TO pgtuner_monitor;
GRANT EXECUTE ON FUNCTION hypopg_unhide_index(oid) TO pgtuner_monitor;
GRANT EXECUTE ON FUNCTION hypopg_relation_size(oid) TO pgtuner_monitor;

-- Note: HypoPG operations are session-scoped and don't affect the actual database

Complete Setup Script

-- 1. Create the monitoring user
CREATE USER pgtuner_monitor WITH PASSWORD 'secure_password';

-- 2. Grant connection and schema access
GRANT CONNECT ON DATABASE your_database TO pgtuner_monitor;
GRANT USAGE ON SCHEMA public TO pgtuner_monitor;

-- 3. Grant read access to user tables
GRANT SELECT ON ALL TABLES IN SCHEMA public TO pgtuner_monitor;
ALTER DEFAULT PRIVILEGES IN SCHEMA public GRANT SELECT ON TABLES TO pgtuner_monitor;

-- 4. Grant system statistics access
GRANT pg_read_all_stats TO pgtuner_monitor;  -- PostgreSQL 10+

-- Grant access to pg_stat_statements views explicitly
GRANT SELECT ON pg_stat_statements TO pgtuner_monitor;
GRANT SELECT ON pg_stat_statements_info TO pgtuner_monitor;

-- 5. Install and grant access to extensions (as superuser)
-- pg_stat_statements (required)
CREATE EXTENSION IF NOT EXISTS pg_stat_statements;

-- pgstattuple (for bloat detection)
CREATE EXTENSION IF NOT EXISTS pgstattuple;
GRANT pg_stat_scan_tables TO pgtuner_monitor;  -- PostgreSQL 14+
-- OR grant individual functions:
-- GRANT EXECUTE ON FUNCTION pgstattuple(regclass) TO pgtuner_monitor;
-- GRANT EXECUTE ON FUNCTION pgstattuple_approx(regclass) TO pgtuner_monitor;
-- GRANT EXECUTE ON FUNCTION pgstatindex(regclass) TO pgtuner_monitor;

-- hypopg (for hypothetical index testing)
CREATE EXTENSION IF NOT EXISTS hypopg;
GRANT SELECT ON hypopg_list_indexes TO pgtuner_monitor;
GRANT SELECT ON hypopg_hidden_indexes TO pgtuner_monitor;
GRANT EXECUTE ON FUNCTION hypopg_create_index(text) TO pgtuner_monitor;
GRANT EXECUTE ON FUNCTION hypopg_drop_index(oid) TO pgtuner_monitor;
GRANT EXECUTE ON FUNCTION hypopg_reset() TO pgtuner_monitor;
GRANT EXECUTE ON FUNCTION hypopg_hide_index(oid) TO pgtuner_monitor;
GRANT EXECUTE ON FUNCTION hypopg_unhide_index(oid) TO pgtuner_monitor;
GRANT EXECUTE ON FUNCTION hypopg_relation_size(oid) TO pgtuner_monitor;

-- 6. Verify permissions
SET ROLE pgtuner_monitor;
SELECT * FROM pg_stat_statements LIMIT 1;
SELECT * FROM pg_stat_activity WHERE pid = pg_backend_pid();
SELECT * FROM pgstattuple('pg_class') LIMIT 1;
SELECT * FROM hypopg_list_indexes();
RESET ROLE;

Excluding Specific Users from Monitoring

You can exclude specific PostgreSQL users from query analysis and monitoring results. This is useful for filtering out monitoring or replication users, system accounts, and internal application service accounts.

Set the PGTUNER_EXCLUDE_USERIDS environment variable with a comma - separated list of user OIDs:

# Exclude user IDs 16384, 16385, and 16386
export PGTUNER_EXCLUDE_USERIDS="16384,16385,16386"

To find the OID for a specific PostgreSQL user:

SELECT usesysid, usename FROM pg_user WHERE usename = 'monitoring_user';

When configured, the following queries are filtered:

• pg_stat_activity queries (filters on usesysid column)
• pg_stat_statements queries (filters on userid column)

This affects tools like get_slow_queries, get_active_queries, analyze_wait_events, check_database_health, and get_index_recommendations.

MCP Client Configuration

Add to your cline_mcp_settings.json or Claude Desktop config:

{
  "mcpServers": {
    "pgtuner_mcp": {
      "command": "python",
      "args": ["-m", "pgtuner_mcp"],
      "env": {
        "DATABASE_URI": "postgresql://user:password@localhost:5432/mydb"
      },
      "disabled": false,
      "autoApprove": []
    }
  }
}

Or Streamable HTTP Mode

{
  "mcpServers": {
    "pgtuner_mcp": {
      "type": "http",
      "url": "http://localhost:8080/mcp"
    }
  }
}

Server Modes

1. Standard MCP Mode (Default)

# Default mode (stdio)
python -m pgtuner_mcp

# Explicitly specify stdio mode
python -m pgtuner_mcp --mode stdio

2. HTTP SSE Mode (Legacy Web Applications)

The SSE (Server - Sent Events) mode provides a web - based transport for MCP communication. It's useful for web applications and clients that need HTTP - based communication.

# Start SSE server on default host/port (0.0.0.0:8080)
python -m pgtuner_mcp --mode sse

# Specify custom host and port
python -m pgtuner_mcp --mode sse --host localhost --port 3000

# Enable debug mode
python -m pgtuner_mcp --mode sse --debug

SSE Endpoints:

MCP Client Configuration for SSE: For MCP clients that support SSE transport (like Claude Desktop or custom clients):

{
  "mcpServers": {
    "pgtuner_mcp": {
      "type": "sse",
      "url": "http://localhost:8080/sse"
    }
  }
}

3. Streamable HTTP Mode (Modern MCP Protocol - Recommended)

The streamable - http mode implements the modern MCP Streamable HTTP protocol with a single /mcp endpoint. It supports both stateful (session - based) and stateless modes.

# Start Streamable HTTP server in stateful mode (default)
python -m pgtuner_mcp --mode streamable-http

# Start in stateless mode (fresh transport per request)
python -m pgtuner_mcp --mode streamable-http --stateless

# Specify custom host and port
python -m pgtuner_mcp --mode streamable-http --host localhost --port 8080

# Enable debug mode
python -m pgtuner_mcp --mode streamable-http --debug

Stateful vs Stateless:

• Stateful (default): Maintains session state across requests using mcp - session - id header. Ideal for long - running interactions.
• Stateless: Creates a fresh transport for each request with no session tracking. Ideal for serverless deployments or simple request/response patterns.

Endpoint: http://{host}:{port}/mcp

Available Tools

⚠️ Important Note

All tools focus exclusively on user/application tables and indexes. System catalog tables (pg_catalog, information_schema, pg_toast) are automatically excluded from all analyses.

Performance Analysis Tools

Index Tuning Tools

Database Health Tools

Bloat Detection Tools (pgstattuple)

Vacuum Monitoring Tools

Tool Parameters

get_slow_queries

• limit: Maximum queries to return (default: 10)
• min_calls: Minimum call count filter (default: 1)
• min_mean_time_ms: Minimum mean (average) execution time in milliseconds filter
• order_by: Sort by mean_time, calls, or rows

analyze_query

• query (required): SQL query to analyze
• analyze: Execute query with EXPLAIN ANALYZE (default: true)
• buffers: Include buffer statistics (default: true)
• format: Output format - json, text, yaml, xml

get_index_recommendations

• workload_queries: Optional list of specific queries to analyze
• max_recommendations: Maximum recommendations (default: 10)
• min_improvement_percent: Minimum improvement threshold (default: 10%)
• include_hypothetical_testing: Test with HypoPG (default: true)
• target_tables: Focus on specific tables

check_database_health

• include_recommendations: Include actionable recommendations (default: true)
• verbose: Include detailed statistics (default: false)

analyze_table_bloat

• table_name: Name of a specific table to analyze (optional)
• schema_name: Schema name (default: public)
• use_approx: Use pgstattuple_approx for faster analysis on large tables (default: false)
• min_table_size_gb: Minimum table size in GB to include in schema - wide scan (default: 5)
• include_toast: Include TOAST table analysis (default: false)

analyze_index_bloat

• index_name: Name of a specific index to analyze (optional)
• table_name: Analyze all indexes on this table (optional)
• schema_name: Schema name (default: public)
• min_index_size_gb: Minimum index size in GB to include (default: 5)
• min_bloat_percent: Only show indexes with bloat above this percentage (default: 20)

get_bloat_summary

• schema_name: Schema to analyze (default: public)
• top_n: Number of top bloated objects to show (default: 10)
• min_size_gb: Minimum object size in GB to include (default: 5)

monitor_vacuum_progress

• action: Action to perform - progress (monitor active vacuum operations), needs_vacuum (find tables needing vacuum), autovacuum_status (review autovacuum configuration), or recent_activity (view recent vacuum history)
• schema_name: Schema to analyze (default: public, used with needs_vacuum action)
• top_n: Number of results to return (default: 20)

analyze_disk_io_patterns

• analysis_type: Type of I/O analysis - all (comprehensive), buffer_pool (cache hit ratios), tables (table I/O patterns), indexes (index I/O patterns), temp_files (temporary file usage), or checkpoints (checkpoint I/O statistics)
• schema_name: Schema to analyze (default: public)
• top_n: Number of top I/O - intensive objects to show (default: 20)
• min_size_gb: Minimum object size in GB to include (default: 1)

MCP Prompts

The server includes pre - defined prompt templates for common tuning workflows:

MCP Resources

Static Resources

• pgtuner://docs/tools - Complete tool documentation
• pgtuner://docs/workflows - Common tuning workflows guide
• pgtuner://docs/prompts - Prompt template documentation

Dynamic Resource Templates

• pgtuner://table/{schema}/{table_name}/stats - Table statistics
• pgtuner://table/{schema}/{table_name}/indexes - Table index information
• pgtuner://query/{query_hash}/stats - Query performance statistics
• pgtuner://settings/{category} - PostgreSQL settings (memory, checkpoint, wal, autovacuum, connections, all)
• pgtuner://health/{check_type} - Health checks (connections, cache, locks, replication, bloat, all)

PostgreSQL Extension Setup

HypoPG Extension

HypoPG enables testing indexes without actually creating them. This is extremely useful for:

• Testing if a proposed index would be used by the query planner
• Comparing execution plans with different index strategies
• Estimating storage requirements before committing

Enable HypoPG in Database

HypoPG enables testing hypothetical indexes without creating them on disk.

-- Create the extension
CREATE EXTENSION IF NOT EXISTS hypopg;

-- Verify installation
SELECT * FROM hypopg_list_indexes();

pg_stat_statements Extension

The pg_stat_statements extension is required for query performance analysis. It tracks planning and execution statistics for all SQL statements executed by a server.

Step 1: Enable the Extension in postgresql.conf

Add the following to your postgresql.conf file:

# Required: Load pg_stat_statements module
shared_preload_libraries = 'pg_stat_statements'

# Required: Enable query identifier computation
compute_query_id = on

# Maximum number of statements tracked (default: 5000)
pg_stat_statements.max = 10000

# Track all statements including nested ones (default: top)
# Options: top, all, none
pg_stat_statements.track = top

# Track utility commands like CREATE, ALTER, DROP (default: on)
pg_stat_statements.track_utility = on

⚠️ Important Note

After modifying shared_preload_libraries, a PostgreSQL server restart is required.

Step 2: Create the Extension in Your Database

-- Connect to your database and create the extension
CREATE EXTENSION IF NOT EXISTS pg_stat_statements;

-- Verify installation
SELECT * FROM pg_stat_statements LIMIT 1;

pgstattuple Extension

The pgstattuple extension is required for bloat detection tools (analyze_table_bloat, analyze_index_bloat, get_bloat_summary). It provides functions to get tuple - level statistics for tables and indexes.

-- Create the extension
CREATE EXTENSION IF NOT EXISTS pgstattuple;

-- Verify installation
SELECT * FROM pgstattuple('pg_class') LIMIT 1;

Performance Impact Considerations

💡 Usage Tip

Use pg_test_timing to measure the timing overhead on your specific system before enabling track_io_timing.

💻 Usage Examples

Basic Usage

Find and Analyze Slow Queries

# Get top 10 slowest queries
slow_queries = await get_slow_queries(limit=10, order_by="total_time")

# Analyze a specific query's execution plan
analysis = await analyze_query(
    query="SELECT * FROM orders WHERE user_id = 123",
    analyze=True,
    buffers=True
)

Get Index Recommendations

# Analyze workload and get recommendations
recommendations = await get_index_recommendations(
    max_recommendations=5,
    min_improvement_percent=20,
    include_hypothetical_testing=True
)

# Recommendations include CREATE INDEX statements
for rec in recommendations["recommendations"]:
    print(rec["create_statement"])

Database Health Check

# Run comprehensive health check
health = await check_database_health(
    include_recommendations=True,
    verbose=True
)

print(f"Health Score: {health['overall_score']}/100")
print(f"Status: {health['status']}")

# Review specific areas
for issue in health["issues"]:
    print(f"{issue}")

Find Unused Indexes

# Find indexes that can be dropped
unused = await find_unused_indexes(
    schema_name="public",
    include_duplicates=True
)

# Get DROP statements
for stmt in unused["recommendations"]:
    print(stmt)

Docker

docker pull  dog830228/pgtuner_mcp

# Streamable HTTP mode (recommended for web applications)
docker run -p 8080:8080 \
  -e DATABASE_URI=postgresql://user:pass@host:5432/db \
  dog830228/pgtuner_mcp --mode streamable-http

# Streamable HTTP stateless mode (for serverless)
docker run -p 8080:8080 \
  -e DATABASE_URI=postgresql://user:pass@host:5432/db \
  dog830228/pgtuner_mcp --mode streamable-http --stateless

# SSE mode (legacy web applications)
docker run -p 8080:8080 \
  -e DATABASE_URI=postgresql://user:pass@host:5432/db \
  dog830228/pgtuner_mcp --mode sse

# stdio mode (for MCP clients like Claude Desktop)
docker run -i \
  -e DATABASE_URI=postgresql://user:pass@host:5432/db \
  dog830228/pgtuner_mcp --mode stdio

🔧 Technical Details

Requirements

• Python: 3.10+
• PostgreSQL: 12+ (recommended: 14+)
• Extensions:
  • pg_stat_statements (required for query analysis)
  • hypopg (optional, for hypothetical index testing)

Dependencies

Core dependencies:

• mcp[cli]>=1.12.0 - Model Context Protocol SDK
• psycopg[binary,pool]>=3.1.0 - PostgreSQL adapter with connection pooling
• pglast>=7.10 - PostgreSQL query parser

Optional (for HTTP modes):

• starlette>=0.27.0 - ASGI framework
• uvicorn>=0.23.0 - ASGI server

🤝 Contributing

Contributions are welcome! Please feel free to submit a Pull Request.