MCP server de pgmcp

pgmcp

pgmcp is a read-only PostgreSQL ops/DBA server for the Model Context Protocol, built on the official modelcontextprotocol/go-sdk. It answers the questions a DBA asks during an incident — which statements are slow, why this plan is slow, which indexes are dead weight, which tables autovacuum has fallen behind on, who is blocking whom, how far the standby is lagging — and it is read-only by construction rather than by convention: a dedicated database role that holds no write privilege, a BEGIN READ ONLY transaction with a statement timeout around every single statement, and a SQL parser guard that rejects anything that is not one SELECT/EXPLAIN/SHOW and every function that could mutate state from inside a read-only transaction. Tested against PostgreSQL 16; requires 13 or later.

Installation

Claude Desktop, one click: download pgmcp_<version>.mcpb from Releases and open it. Claude Desktop asks for a Postgres connection string, keeps it in the OS keychain, and launches the bundled binary itself — nothing on PATH, no config file to edit. One bundle covers macOS (universal) and Windows (x64).

Otherwise download a binary for your platform from Releases — darwin, linux and windows, amd64 and arm64, with checksums.

Or build from source with the Go CLI tool go:

go install github.com/pascalallen/pgmcp/cmd/pgmcp@latest

Or run the released image, which is distroless, non-root and multi-arch:

docker run --rm -i -e PGMCP_DATABASE_URL='postgres://…' ghcr.io/pascalallen/pgmcp

pgmcp is listed in the MCP Registry as io.github.pascalallen/pgmcp.

Before you point it at anything, create the read-only role — see Database role. It is the layer that still holds if the other two have a bug.

Usage

One MCP surface, two transports. Which one you run is configuration, not a different build.

Claude Code, stdio — the client launches the binary and talks over stdin/stdout:

claude mcp add pgmcp --transport stdio \
  --env PGMCP_DATABASE_URL='postgres://pgmcp:…@db.internal:5432/app?sslmode=require' \
  -- pgmcp

Claude Desktop, stdio — install the .mcpb bundle from Releases (see Installation), or write the same thing into claude_desktop_config.json by hand:

{
  "mcpServers": {
    "pgmcp": {
      "command": "pgmcp",
      "env": {
        "PGMCP_DATABASE_URL": "postgres://pgmcp:…@db.internal:5432/app?sslmode=require"
      }
    }
  }
}

HTTP — Streamable HTTP behind a static bearer key, for a shared deployment. pgmcp speaks plain HTTP and never terminates TLS itself; run it on loopback behind a reverse proxy.

PGMCP_DATABASE_URL='postgres://pgmcp:…@db.internal:5432/app?sslmode=require' \
PGMCP_AUTH_MODE=static \
PGMCP_API_KEYS="$(openssl rand -hex 32)" \
  pgmcp --transport http --listen 127.0.0.1:8080

claude mcp add pgmcp --transport http https://pgmcp.example.com/mcp \
  --header "Authorization: Bearer <key>"

TLS termination, the proxy settings the streaming transport needs, JWT auth against an identity provider, and wiring pgmcp up as a claude.ai custom connector are all in docs/DEPLOYING.md.

Tools

Tool	The question it answers
top_queries	Which statements are slow or expensive server-wide? Ranks pg_stat_statements by total time, mean time, calls, rows or blocks read.
explain	Why is this statement slow? Plan tree, the nodes burning the most self time, plan warnings, and a stable plan_hash to diff against a later run.
index_health	Which indexes can I drop, and which are missing their job? Never-scanned, duplicate, invalid and bloated indexes.
table_health	Where is autovacuum falling behind? Dead-tuple ratio, last vacuum/analyze, sequential versus index scans and estimated bloat, per table.
lock_waits	Why is this query hanging? The current lock wait graph — who is blocked, who blocks them, and any cycle that amounts to a deadlock.
connections	What is the server doing right now, and how close is it to max_connections? Backends grouped by state, wait event, application, user or database, including idle-in-transaction sessions.
replication	How far behind is the standby, and which slot is retaining WAL? Primary/standby role, per-standby lag in bytes and milliseconds, slots and the current WAL rate.
config_check	Is this server tuned sanely? pg_settings against memory, autovacuum, WAL and connection heuristics, with an ok/review/warn verdict and a note per setting.
query	Everything the other eight do not cover. One read-only SELECT/EXPLAIN/SHOW in a READ ONLY transaction, bounded by a row cap and a statement timeout, with $1..$n bind parameters.

Every tool is annotated readOnlyHint: true, destructiveHint: false, idempotentHint: true, openWorldHint: false, and returns a typed output schema.

query is the only tool that carries free-form SQL, and it is optional. --disable-query drops it from the catalogue entirely — a deployment that only needs the eight diagnostics tools can run without any ad hoc SQL surface at all. --query-schemas=public,app restricts it to named schemas instead — and bounds explain with it, since analyze=true executes the statement; read what that does and does not stop in docs/SECURITY.md.

Resources & prompt

Resource	Contents
pgmcp://overview	The server snapshot to start from: version, uptime, recovery state, installed extensions, per-database sizes, cache hit ratio, and connections against max_connections. Cacheable for 30s.
pgmcp://settings	The raw pg_settings rows. Cacheable for 5m.

Prompt	Arguments	Purpose
diagnose_slow_query	sql (required)	A four-step investigation: explain the statement, check index_health/table_health on every schema in the plan's hot nodes, look for it in top_queries, then summarize root cause, evidence and a recommended index or rewrite — written out as text only, never run.

Configuration

Every setting has a --flag and a PGMCP_<KEY> environment variable. Flags win over the environment, which wins over the default. Configuration errors exit 2 with every offending key named in one message; runtime failures exit 1.

Flag	Env	Default	Meaning
--database-url	PGMCP_DATABASE_URL	— (required)	Postgres connection string
--transport	PGMCP_TRANSPORT	stdio	stdio or http
--listen	PGMCP_LISTEN	127.0.0.1:8080	HTTP listen address
--resource-url	PGMCP_RESOURCE_URL	—	Public origin this server is reachable at, for OAuth resource metadata
--auth-mode	PGMCP_AUTH_MODE	none	none, static or jwt (HTTP only)
--api-keys	PGMCP_API_KEYS	—	Comma-separated static API keys, required for static
--jwks-url	PGMCP_JWKS_URL	—	JWK set URL, required for jwt
--jwt-issuer	PGMCP_JWT_ISSUER	—	Required iss claim, required for jwt
--jwt-audience	PGMCP_JWT_AUDIENCE	—	Required aud claim, required for jwt
--auth-servers	PGMCP_AUTH_SERVERS	—	Comma-separated OAuth authorization servers to advertise via RFC 9728
--disable-query	PGMCP_DISABLE_QUERY	false	Drop the ad hoc query tool entirely
--query-schemas	PGMCP_QUERY_SCHEMAS	—	Comma-separated schemas the query and explain tools may read; unset disables the allowlist
--max-conns	PGMCP_MAX_CONNS	4	Maximum Postgres connections
--call-timeout	PGMCP_CALL_TIMEOUT	60s	Per-tool-call timeout
--rate-limit	PGMCP_RATE_LIMIT	60	Tool calls per principal per minute (HTTP only)
--max-output-bytes	PGMCP_MAX_OUTPUT_BYTES	1048576	Cap on a tool call's structured content
--log-level	PGMCP_LOG_LEVEL	info	debug, info, warn or error
--log-format	PGMCP_LOG_FORMAT	text	text or json
--insecure-no-auth	PGMCP_INSECURE_NO_AUTH	false	Allow auth-mode=none on a non-loopback listen address
--version	—	—	Print the version and exit

The auth block applies to the HTTP transport only. Over stdio the operating system decides who the caller is: the parent process that launched the binary, and nobody else.

Security model

• Read-only three independent ways. A dedicated role with no write privilege (pg_monitor plus SELECT, and deliberately not pg_signal_backend); BEGIN READ ONLY with SET LOCAL statement_timeout and lock_timeout = '2s' around every statement the adapter runs, always rolled back; and a parser-level guard, because a read-only transaction alone does not stop pg_terminate_backend, pg_read_file, pg_sleep or setval.
• The SQL guard is allow-list first. One top-level statement, and it must be a SELECT, EXPLAIN or SHOW; no nested write statement anywhere in the tree; no FOR UPDATE/FOR SHARE locking clause; no SELECT INTO; and no call to a denied function — file access, backup and WAL control, replication slots, advisory locks, dblink, sequence mutation, stats resets.
• The schema allowlist is a guardrail, not a boundary. --query-schemas matches the schemas qualifying table references in the parsed statement, case-insensitively, and it bounds both tools that carry caller-supplied SQL — query and explain, so explain with analyze=true cannot execute against a schema you excluded. A view, a set-returning function, or a SECURITY DEFINER function inside an allowed schema can still read outside it. Database privileges are the boundary; the allowlist just narrows the obvious path.
• Authenticated, fail-closed, over HTTP. Static keys are compared in constant time against every stored hash without an early exit; JWTs are validated against a JWK set with asymmetric algorithms only (no alg=none, no HMAC confusion) and a required iss, aud and exp, and the verifier holds no keys until the JWKS arrives, so it starts closed rather than open. RFC 9728 protected resource metadata advertises where to get a token. The server refuses to start on a non-loopback address with auth off.
• Bounded. Per-principal rate limiting, a per-call timeout, a statement timeout and lock timeout inside the transaction, a row cap on the query tool, a cap on a result's structured content, and a 1 MiB request body limit.
• Nothing sensitive is logged. A tool call logs its name, duration, outcome and the caller's user id — never arguments, SQL text, result rows or error text. Parse failures come back as a fixed phrase rather than echoing the statement, and the DSN is redacted from connection errors.

The threat model, the full enumeration of the layers, and the limitations each one does not cover are in docs/SECURITY.md.

Testing

Run the test suite with the race detector and coverage:

go test -race -cover ./...

Integration tests need a Postgres database and are skipped when PGMCP_TEST_DSN is unset. To run them against a scratch Postgres with pg_stat_statements preloaded:

docker run -d --rm --name pg -e POSTGRES_PASSWORD=postgres -p 5544:5432 postgres:16 \
  -c shared_preload_libraries=pg_stat_statements -c pg_stat_statements.track=all
docker exec pg psql -U postgres -c "CREATE EXTENSION IF NOT EXISTS pg_stat_statements"
PGMCP_TEST_DSN="postgres://postgres:postgres@localhost:5544/postgres?sslmode=disable" go test -race -cover ./...

Create and view a coverage profile:

go test -covermode=count -coverprofile=coverage.out ./...
go tool cover -html=coverage.out

Drive a running server through the official MCP conformance suite, or smoke it with the Inspector:

npx -y @modelcontextprotocol/conformance server --url http://127.0.0.1:8080/mcp \
  --expected-failures .github/conformance-expected-failures.yaml
npx @modelcontextprotocol/inspector --cli http://127.0.0.1:8080/mcp --transport http --method tools/list

Contributing

Pull requests are welcome. For major changes, please open an issue first to discuss what you would like to change.

Please make sure to update tests as appropriate.

License

MIT