PostgreSQL Performance (read-only)

pg-perf-mcp

An MCP server that lets AI assistants diagnose PostgreSQL performance problems safely: slow queries, EXPLAIN plans, index suggestions, table health and lock contention. It is read-only by design.

*A real Claude Code session against the demo shop, sped up. First it finds which database functions are slow and why. Then it explains a slow catalog query and estimates the fix with a hypothetical index. Asked to create the index, it says it can't: every tool is read-only.*

Why

When a page is slow, developers paste a query into a chat and ask "why is this slow?". The assistant then guesses, because it cannot see the plan, the table sizes, the existing indexes or which queries actually dominate the load.

pg-perf-mcp gives the assistant real numbers from `pg_stat_statements`, `EXPLAIN` and the catalogs, through a narrow, audited, read-only interface. It cannot change data, cannot run DDL, and never creates the indexes it recommends.

Demo

A condensed, illustrative session against the demo shop (docker compose up, then ./demo/workload.sh). Exact numbers vary from run to run:

You:    The shop database feels slow. What's going on?

Claude: [top_slow_queries orderBy=total_time] The top statement by total time is SELECT id, status, total, created_at FROM shop.orders WHERE customer_id = $1 ORDER BY created_at DESC (thousands of calls, the largest share of total execution time).

[explain_query sql="... WHERE customer_id = 42 ORDER BY created_at DESC"] HIGH SEQ_SCAN_ON_LARGE_TABLE Seq Scan on shop.orders reads all ~1.0M rows of shop.orders to apply filter (customer_id = 42); only a handful match.

[suggest_indexes] CREATE INDEX CONCURRENTLY idx_orders_customer_id ON shop.orders (customer_id); hypopg: the planner would use it; estimated cost drops by well over 90%. Not executed. Review it and run it yourself, ideally off-peak.

[table_health table=shop.orders] Autovacuum is disabled on shop.orders and ~20% of rows are dead. Re-enable autovacuum and VACUUM.

Prioritized actions: 1) add the index above (orders.customer_id is a foreign key with no index),

1. re-enable autovacuum on shop.orders, 3) ...

Architecture

flowchart LR

A["AI assistant (Claude Code, Claude Desktop, ...)"] -- "MCP over streamable HTTP Authorization: Bearer key" --> F["API key filter (constant-time)"] F --> M["MCP tools / resources / prompt Spring AI 2.0 annotations"] M --> R["ToolRunner audit log · metrics · safe errors"] R --> G["SqlGuard PG lexer + JSqlParser"] G --> E["ReadOnlyExecutor READ ONLY tx · always rollback statement/lock timeout · row cap"] R --> AN["analysis PlanAnalyzer · IndexAdvisor (pure functions)"] E --> P[("PostgreSQL 18 role mcp_readonly pg_stat_statements · hypopg")]

Tools

Resources: pg://schema/overview (tables, row estimates, sizes, indexes) and pg://schema/{table} (columns, indexes with scan counts, foreign keys flagged when unindexed). Prompt: diagnose_slow_database, a guided workflow: slow queries, then explain, suggest indexes, table health, functions, then a prioritized summary.

Why a separate tool for database functions

EXPLAIN shows a function call as one opaque expression, and charges its time to whichever plan node evaluates it. On the demo catalog page (20 products, each calling shop.product_rating), explain_query blames the index scan on products (184 ms self time) and finds nothing else serious. explain_function on the same query, against the full demo data:

shop.product_rating          plpgsql  20 calls  99% of execution time

SELECT avg(rating) FROM shop.reviews WHERE product_id = p_product_id 20 calls 175 ms 96% of execution time HIGH FUNCTION_DOMINATES_QUERY shop.product_rating takes 99% of the query's execution time HIGH FUNCTION_CALLED_PER_ROW ran 20 times for one query (9.0 ms per call) HIGH SEQ_SCAN_ON_LARGE_TABLE inside shop.product_rating: Seq Scan on shop.reviews reads all ~250k rows to apply filter (reviews.product_id = $1)

How it works (ADR 0006): the query runs once with auto_explain enabled for that transaction only, so the plan of each nested statement comes back as a notice; per-function time comes from pg_stat_xact_user_functions and per-statement calls from pg_stat_statements, both as before/after differences.

Security model

Five independent layers stand between the assistant and a write (details in ADR 0001):

1. SQL guard: exactly one plain SELECT. It is checked by a PostgreSQL-faithful lexer *and* JSqlParser, and both must agree. Blocks DML/DDL, multiple statements, data-modifying CTEs, SELECT INTO, FOR UPDATE/SHARE, COPY, DO, and dangerous functions (pg_sleep*, *_file, lo_*, dblink*, set_config, pg_terminate_backend, query_to_xml, ...). It fails closed.
2. Read-only role: mcp_readonly has SELECT plus pg_read_all_stats/pg_monitor, with default_transaction_read_only = on and statement_timeout = 5s set on the role.
3. Read-only connections: Hikari readOnly=true and a read-only session default.
4. One rolled-back read-only transaction per call: statement_timeout/lock_timeout set locally, standard_conforming_strings forced on, always rolled back.
5. Row cap: every result is capped (default 200 rows).

Around those layers: a bearer API key (constant-time comparison; only /actuator/health is open), a JSON audit line per call (SQL stored as a SHA-256 hash plus an 80-character preview), error messages with no stack traces or connection details, and metrics mcp.tool.duration and mcp.guard.rejections.

Quickstart

Requirements: Docker. There are two ways to start:

• [Use it on your database](#use-it-on-your-database): pull the published image. No clone or build.
• [Try the demo shop](#try-the-demo-shop): clone the repo and get a seeded database with realistic problems to diagnose.

Either way, finish with connecting your AI client.

Use it on your database

1. Pull the image (public on GHCR, built for amd64 and arm64):

docker pull ghcr.io/gouranshul/pg-perf-mcp:0.1.0

2. Prepare the database (once). Create a read-only role like mcp_readonly (see docker/init.sql) and make sure pg_stat_statements is in shared_preload_libraries. Prefer a replica or staging database: analyze=true and explain_function execute the query. For the function tools, also enable function tracking.

3. Run it. Put the settings in a file so secrets stay out of your shell history:

cat > pg-perf.env <<'EOF'

PGPERF_DB_URL=jdbc:postgresql://host.docker.internal:5432/mydb PGPERF_DB_USER=mcp_readonly PGPERF_DB_PASSWORD= MCP_API_KEY=<32+ random characters> EOF

docker run --rm --env-file pg-perf.env -p 127.0.0.1:8080:8080 ghcr.io/gouranshul/pg-perf-mcp:0.1.0

host.docker.internal reaches a database on your machine (Docker Desktop; on Linux add --add-host=host.docker.internal:host-gateway). The MCP endpoint is now http://localhost:8080/mcp.

4. Connect your AI client: see below.

Try the demo shop

To build and test from source you also need Java 25.

git clone https://github.com/gouranshul/pg-perf-mcp && cd pg-perf-mcp

cp .env.example .env # then replace the placeholder secrets docker compose up --build # Postgres 18 + seeded demo shop (~4.5M rows) + the MCP server on :8080

Generate some slow-query statistics (optional, but makes top_slow_queries interesting):

./demo/workload.sh          # 60s of deliberately bad queries via pgbench

./demo/lock-scenario.sh # holds a row lock for 2 minutes so blocking_sessions has something to show ./demo/smoke-test.sh # curl-based end-to-end check of health, auth and two tool calls

Connect from Claude Code

claude mcp add --transport http pg-perf http://localhost:8080/mcp --header "Authorization: Bearer $MCP_API_KEY"

In PowerShell, set the key first, or the header goes out empty and the server answers 401:

$env:MCP_API_KEY = "<your MCP_API_KEY>"

claude mcp add --transport http pg-perf http://localhost:8080/mcp --header "Authorization: Bearer $env:MCP_API_KEY"

Or install it as a Claude Code plugin. Claude Code asks for the endpoint (default http://localhost:8080/mcp) and your MCP_API_KEY, which it keeps in secure storage:

/plugin install pg-perf --marketplace gouranshul/pg-perf-mcp

On Claude Code versions before 2.1.275, add the marketplace first: `/plu