pgc

pgc compiles annotated SQL queries into a type-safe Go package for PostgreSQL. You write plain SQL; pgc asks your development database what every parameter and column really is - the server itself is the type oracle, no SQL parsing involved - and generates code that reads like a person wrote it.

The generated package imports only the standard library (database/sql, context, time, ...). pgc itself has zero third-party dependencies: it speaks the PostgreSQL wire protocol directly.

Installation

go install github.com/goloop/pgc@latest

pgc is stable from v1: the commands and their exit codes, pgc.json, the pgc.lock.json format, the migration history table, the connection URL and the shape of the generated code do not change incompatibly within v1. The "Compatibility" section of DOC.md spells out what that covers. Pin the version in CI (@v1.0.1 rather than @latest).

Building from source requires Go 1.24 or newer. Prebuilt binaries for Linux, macOS and Windows are attached to the releases, with checksums.

Either way, generation needs a PostgreSQL database to ask (a disposable local container is perfect).

Quick start

Write queries next to your migrations:

-- queries/users.sql

-- name: GetUser :one
-- Returns a single user by primary key.
SELECT id, email, name, created_at
FROM users
WHERE id = $1;

-- name: CreateUser :one
INSERT INTO users (email, name)
VALUES (@email, @name)
RETURNING id, email, name, created_at;

Parameters are $1..$N or, when a statement has enough of them that counting becomes the risk, @name - numbered for you, and named after what you wrote.

Point pgc at your development database and generate:

export PGC_DATABASE_URL="postgres://user:pass@localhost:5432/app?sslmode=disable"
pgc generate

Use the result like hand-written code, because it looks like hand-written code:

q := db.New(sqlDB) // *sql.DB

u, err := q.CreateUser(ctx, "ada@example.com", "Ada")
got, err := q.GetUser(ctx, u.ID)

// Inside a transaction - same methods:
tx, _ := sqlDB.BeginTx(ctx, nil)
err = q.WithTx(tx).DeleteUser(ctx, u.ID)

Every parameter and result is typed from the live schema: bigint becomes int64, a nullable text becomes *string, timestamptz becomes time.Time, a PostgreSQL enum becomes a named string type with constants, text[] becomes []string (including WHERE id = ANY($1)), and a domain resolves to its base type. When the full column set of a table is selected, the query returns the table's model struct; projections get their own row struct; a single column comes back as a bare value; a joined table can be nested with -- embed:. Expressions that cannot produce NULL - count(...), coalesce(..., <literal>) - come back as plain values rather than pointers. Optional extras: json:"column" tags, a Querier interface for test doubles, and initialisms for your domain's abbreviations, so seo_title becomes SEOTitle instead of SeoTitle.

Commands

pgc generate [-c pgc.json] [-d url]  compile the queries into a Go package
pgc check    [-c pgc.json] [-d url]  fail unless the package is up to date
pgc verify   [-c pgc.json] [-d url]  check the query types in pgc.lock.json
                                     against the database
pgc migrate  [up] [-c pgc.json] [-d url]
                                     apply pending migrations, in order
pgc migrate status                   list every migration's state
pgc migrate resolve <file> applied|retry
pgc migrate baseline <last-file>
pgc describe [-d url] "SELECT ..."   print parameter and column types
pgc version                          print the version

pgc migrate applies plain-SQL migration files exactly once each, in name order - one transaction per file, an advisory lock against concurrent runs and a history table checked before anything runs, so the whole database lifecycle lives in one tool: pgc migrate, pgc generate, pgc check. An applied file that was edited or deleted stops the run, a file with its own COMMIT is refused, and a no-transaction file that fails half way is never silently repeated. In CI, pgc check fails when the generated package no longer matches its queries, and pgc migrate status when the migration history needs attention.

Generating needs a database only once. pgc generate records what the server said in pgc.lock.json; commit it, and every later run without a database URL generates the identical package from that record - a fresh clone, a CI job, a container build. The record is refused the moment a query stops matching it, and pgc verify reports drift in the one job that does have a database.

Documentation

New to pgc? The step-by-step TUTORIAL.md (Ukrainian: TUTORIAL.UK.md) takes you from an empty directory to a running program in about twenty minutes - Docker database, migrations, annotations, generation and usage, all copy-pasteable.

Full reference: DOC.md (Ukrainian: DOC.UK.md) - migrations, annotations, configuration, the type mapping, nullability rules, the anatomy of the generated code and the compatibility promise.

Changes between releases: CHANGELOG.md.

Contributing

Before submitting changes, run:

gofmt -l .
go vet ./...
go test ./...

The integration tests run against a real PostgreSQL. They create and drop databases of their own, so point them at a disposable server only:

docker run -d --rm --name pgc-test -e POSTGRES_PASSWORD=test \
  -p 127.0.0.1:55432:5432 postgres:17-alpine
PGC_DATABASE_URL="postgres://postgres:test@127.0.0.1:55432/postgres?sslmode=disable" \
  go test -tags integration ./...

License

MIT - see LICENSE.

Release v1.0.2

Bound query result storage and timeouts, and build SQL without repeated copies.