Write SQL. Keep TypeScript. Skip the query-builder translation layer.

SQLBraid is a SQL-first data-access toolkit for TypeScript. It keeps ordinary SQL visible while adding safe value binds, readable dynamic SQL, explicit result contracts, Standard Schema result mapping, physical connection ownership, and driver-owned transports.

On Node.js 22.18 or newer, save this as quickstart.mts and run node quickstart.mts. It creates and closes its own in-memory database:

import { createNodeSqliteDatabase, sql } from "sqlbraid/node-sqlite";
import { DatabaseSync } from "node:sqlite";

interface UserRow {
  id: string;
  name: string;
}

const native = new DatabaseSync(":memory:");
try {
  native.exec("CREATE TABLE users (id INTEGER PRIMARY KEY, name TEXT NOT NULL)");
  native.prepare("INSERT INTO users (name) VALUES (?)").run("Ada");
  const db = createNodeSqliteDatabase(native);
  const userId = 1;
  const users = await db.all(sql.rows<UserRow>`
    SELECT id, name FROM users WHERE id = ${userId}
  `);
  console.log(users); // [{ id: "1", name: "Ada" }]
} finally {
  native.close();
}

Application code installs the unscoped sqlbraid facade and imports a combined driver+dialect/query subpath such as sqlbraid/pg, sqlbraid/mysql2, sqlbraid/mariadb, sqlbraid/node-sqlite, sqlbraid/better-sqlite3, sqlbraid/libsql, sqlbraid/sqlite-wasm, sqlbraid/d1, sqlbraid/oracledb, or sqlbraid/tedious. The facade has no implicit default dialect; its root exports only common runtime contracts. The granular @sqlbraid/* packages remain available for custom integrations and tooling.

Release status: 1.0.0 — stable public API. GA means stable public contracts, not universal driver capabilities. Versioned support records identify each certified database/driver/profile/runtime tuple, implementation revision, and workflow evidence. Changed revisions require fresh exact-SHA Runtime, Documentation, and Release gates; neighboring versions do not inherit certification. No tag, npm publication, Pages deployment, or release authorization is implied.

Get started · Documentation · Architecture mental model · Data representations · Public API audit · Driver-author guide

• Ordinary ${value} interpolation is always a value bind.
• Structural SQL uses explicit helpers such as sql.ident, sql.fragment, sql.list, sql.join, sql.raw, and sql.empty.
• @braid directives (if, choose, when, otherwise, where, set, trim) are lowered by the compiler; inactive branches stay lazy.
• The renderer produces one immutable logical RenderedStatement: segments.length === parameters.length + 1. A rendered parameter is never SQL, an identifier, a nested query, or a driver fragment.
• The selected adapter owns placeholder materialization. $1, ?, :1, @p1, and native value-template syntax are transport details, not logical shape identity.

const query = sql.rows<UserRow>`
  SELECT id, name
  FROM users
  /*@braid where*/
    /*@braid if ${teamId != null}*/
      AND team_id = ${teamId}
    /*@braid end*/
  /*@braid end*/
`;

Use sql.bind(value, hint) only when a first-party adapter documents the database parameter metadata. Hints are not application codecs or Standard Schema validators; an adapter honors a hint or rejects it before I/O.

sql.rows<UserRow>`SELECT ...`;
sql.command`UPDATE ...`;
sql.call({ resultSets: [UserSchema] as const })`CALL ...`;
sql`driver-specific SQL`; // unknown result kind

db.all, db.one, db.maybeOne, and db.stream require sql.rows. db.execute accepts row, command, or unknown queries and checks the actual result kind after execution. db.call accepts sql.call and returns output, ordered heterogeneous resultSets, and an optional returnValue.

sql.out(name, hint?) is valid for sql.call and Oracle row-returning DML. sql.inOut(name, value, hint?) is call-only. Cursor and emitted result sets are materialized and closed before asynchronous mapping; raw cursors, portals, requests, and carrier rows do not escape. A result-kind mismatch is BRAID_RESULT_KIND after execution and cannot undo a root side effect.

Routine channels depend on the adapter: mysql2 supports emitted CALL result sets, but not OUT/INOUT descriptor carriers; SQLite adapters do not support db.call (routine.call), which is an API capability limit, not a restriction on authored SQLite SQL. PostgreSQL refcursors require an existing db.tx; direct SQL Server cursor OUT is unsupported. Unsupported routine channels fail explicitly rather than being emulated; see the support records. callStream is reserved and unimplemented, not a callable API.

Attach a Standard Schema to a row query or pass one per execution:

const eventQuery = sql.rows(EventSchema)`SELECT created_at, payload FROM events`;
const event = await db.one(eventQuery, { schema: EventSchema });

Mapping is one row to one application value. SQLBraid does not hydrate relations, maintain identity maps, or infer arbitrary SELECT/JOIN result types.

The public runtime surface is intentionally small:

interface ExecutionOptions { signal?: AbortSignal }
interface RowValidationOptions<Row> extends ExecutionOptions { schema?: StandardSchemaV1<unknown, Row> }
interface StreamOptions<Row> extends RowValidationOptions<Row> {}

type TransactionIsolation =
  | "read-uncommitted" | "read-committed" | "repeatable-read" | "serializable";
interface TransactionOptions {
  isolation?: TransactionIsolation;
  readOnly?: boolean;
}

await db.execute(query, options?);
await db.all(rows, options?);
await db.one(rows, options?);
await db.maybeOne(rows, options?);
await db.call(call, options?);
await db.batch(queries, options?);
await db.bulk(inputs, factory, options?);
await db.environment(options?);
db.stream(rows, options?);
await db.session(callback);
await db.tx(callback);
await db.tx(transactionOptions, callback);

All execution methods accept options in the trailing position. signal is an AbortSignal; cancellation is capability-driven. An already-aborted signal rejects with its reason. An active signal requires the adapter's statement.cancel capability; otherwise the operation fails with UnsupportedFeatureError (BRAID_CANCEL_UNSUPPORTED) rather than pretending cancellation is supported.

A direct database wraps one physical executor. A pooled database wraps a ConnectionProvider:

interface ConnectionProvider {
  readonly statementBinding: StatementBindingAdapter;
  acquire(): Promise<ConnectionLease>;
}

interface ConnectionLease extends QueryExecutor {
  release(options?: { discard?: boolean }): void | Promise<void>;
}

A provider is a source of leases; it is not itself a physical connection and must not be modeled as a fake executor whose transaction commands can land on unrelated connections. Each pooled root operation acquires one lease, performs physical I/O, releases it, and then maps materialized results. A stream keeps its lease until the driver resource closes. The application owns pool shutdown.

db.session(async (session) => ...) acquires one lease for the callback and reuses that physical lease for nested operations and nested sessions. db.tx(...) inside a session uses the session lease; it does not reacquire. The root database must not be used to escape the session. An unavailable session primitive fails with BRAID_SESSION_UNSUPPORTED; root misuse, closed callback handles, and sibling/parent transaction handles fail with the runtime's scope errors.

db.tx begins and ends one physical transaction. If transactions are absent, the operation fails with BRAID_TX_UNSUPPORTED. Nested transactions use savepoints when the executor exposes them. Explicit transaction options are supported only when the adapter advertises the matching capability; valid but unsupported isolation/access options fail with UnsupportedFeatureError (BRAID_TX_OPTION_UNSUPPORTED, feature transaction.isolation.<level> or transaction.read-only). Malformed runtime values fail before acquisition with TypeError / BRAID_TX_OPTIONS_INVALID. Nested tx(options, callback) is rejected with BRAID_TX_OPTIONS_NESTED rather than silently changing an active transaction. If options are omitted, the database/driver default remains in force; SQLBraid does not guess a profile or reset a session.

await db.tx({ isolation: "serializable", readOnly: true }, async (tx) => {
  await tx.all(sql.rows<{ id: string }>`SELECT id FROM accounts`);
});

Transaction-control uncertainty poisons a direct resource or discards a pooled lease. Use the innermost callback handle while a savepoint is active.

A successful callback is not sufficient for transaction success: PostgreSQL and Bun PostgreSQL reject with BRAID_TX_NOT_COMMITTED if the server reports that COMMIT actually rolled back.

prepare accepts an input factory (required input by default, or explicitly { input: "required" }) or a zero-input factory that explicitly declares { input: "none" }. The factory is evaluated per execution, renders once, and is locked to the first logical shape: result kind, canonical segments, ordered hint/direction/output metadata, and dialect. Values may change; shape changes fail before driver I/O with BRAID_PREPARED_SHAPE.

const byId = db.prepare(
  "user-by-id",
  (id: string) => sql.rows<UserRow>`
  SELECT id, name FROM users WHERE id = ${id}
`,
);

await byId.execute("u_1", { signal });
await byId.all("u_1", { schema: UserSchema });
await byId.one("u_1");
await byId.maybeOne("u_1");
for await (const row of byId.stream("u_1", { signal })) console.log(row);

A zero-input prepared query is declared with db.prepare("users", () => query, { input: "none" }) and takes options as its only execution argument (prepared.all({ signal })). Row queries expose execute, all, one, maybeOne, and stream; command/unknown queries expose execute; call queries expose call. Prepared means a stable SQLBraid application shape, not a universal native/server prepared cache. The adapter reports effective reuse.

DatabaseOptions accepts observers. Events include query:ready, query:result, query:mapped, query:error, bulk:ready, bulk:result, stream:start, stream:end, and transaction. query:ready is emitted after pure binding description and before lease acquisition. It exposes the effective adapter/dialect/transport/reuse plan, declared kind, values and lazy literalizedSql() diagnostics. Bind values are not logged by SQLBraid; applications own redaction and retention.

Observers run in registration order and are observe/fail-only: they may inspect or throw, but cannot rewrite SQL, replace binds/results, retry, or route. A pre-I/O failure prevents execution; a post-I/O failure cannot undo a root side effect. Timing fields are named durationMs.

Optional @sqlbraid/opentelemetry adds SQLBraid-level DB client spans and the stable db.client.operation.duration metric. It keeps SDK/exporter ownership with the application, omits bind values and literalized SQL, and does not add itself to the sqlbraid facade.

These are independent axes:

dialect  SQL surface, quoting, lexical behavior
 driver  protocol/API bridge, binding/materialization, result normalization
runtime  Node, Bun, Deno, browser, or Worker host

First-party dialect roots are PostgreSQL, MySQL, MariaDB, SQLite, Oracle, and SQL Server. Driver subpaths include pg, mysql2, MariaDB Connector/Node.js, node:sqlite, better-sqlite3, libSQL, SQLite WASM, D1, node-oracledb Thin, and Tedious. A new JavaScript runtime does not require a new dialect or adapter when an existing driver API works.

SQLite keeps one dialect while exposing driver-specific subpaths: sqlbraid/node-sqlite, sqlbraid/better-sqlite3, sqlbraid/libsql, sqlbraid/sqlite-wasm, and sqlbraid/d1. Node node:sqlite and better-sqlite3 execute synchronously at the physical boundary; the public Database remains async, and the Awaitable<T> SPI type avoids adding unnecessary Promise wrappers. Synchronous better-sqlite3 calls still block the JavaScript event loop. libSQL requires an explicit { intMode: "string" } assertion for exact INTEGER strings, uses its interactive transaction handle for transaction continuity, does not claim pinned ordinary sessions, and rejects streaming rather than buffering.

Bun's first-party SQL adapter is a single adapter family. The user selects dialect: "postgres" | "mysql" | "mariadb" | "sqlite"; the adapter does not auto-detect SQL semantics from a connection. Bun 1.3.14 has no supported active cancellation (BRAID_CANCEL_UNSUPPORTED) and stream/routine carriers remain unsupported. Its result.rows/result.command metadata is guarded by bun-sql.result-kind-metadata; for Bun MySQL/MariaDB, an empty SELECT and zero-affected DML/DDL are BRAID_RESULT_KIND_AMBIGUOUS after execution because the driver reports command: null and affectedRows: 0, so side effects may already have occurred. Nonempty rows and positive command counts are the supported cases. Deno uses existing first-party driver adapters where their public Node-compatible API works; it does not receive a new Deno-specific dialect. Runtime labels are evidence labels, not promises: Official requires the exact runtime/driver/database/profile tuple in support evidence; otherwise use Compatible, Custom, or Unsupported.

Bun.SQL MySQL/MariaDB do not support explicit transaction access modes: both readOnly: true and readOnly: false reject before I/O with BRAID_TX_OPTION_UNSUPPORTED (transaction.read-only). Bun 1.3.14 can retain a read-only statement failure after rollback on the same connection. Omitting readOnly preserves the native session default; it does not force read-write. This restriction does not apply to Bun.SQL PostgreSQL.

Do not promote an unverified server version, runtime, profile option, or capability from a neighboring tuple. db.environment({ targets?, refresh? }) is observational and returns compatible when no single verified exact target matches. Capability keys use the canonical names:

statement.prepare       statement.stream       statement.bulk
transaction             transaction.savepoint
routine.out             routine.result-sets    routine.out-cursor
routine.return-value

Support capability is capability-driven rather than a promise that every dialect combination has every operation. DML RETURNING/OUTPUT remains native SQL and materialized unless the selected adapter's evidence says otherwise.

SQLBraid is not an ORM, query-builder-first language, complete SQL parser, universal SQL compiler, input-codec framework, retry/router, audit-log store, or connection-pool implementation. It does not rewrite SQL, infer arbitrary result models, hydrate object graphs, or silently emulate missing cursors, transactions, cancellation, or routine channels. Metadata is open-world positive evidence; missing facts do not prove SQL invalid. Numeric exactness, JSON/temporal profiles, and nested containers remain separate evidence boundaries. Exact database numerics are canonical strings; approximate IEEE values are numbers; decodeExactInteger is an application opt-in transform. Bun's 1.3.14 profile uses { bigint: true } for PostgreSQL/MySQL/MariaDB and { safeIntegers: true } for SQLite. Integral Number rows reject; PostgreSQL decimal is text, MySQL/MariaDB DECIMAL and binary byte carriers require authored SQL text/hex conversion, and SQLite native decimal is unsupported. Bun SQLite's mixed-quote SQL classifier also limits result-kind guarantees; bind JSON values. null is SQL NULL; ordinary undefined IN binds fail before acquisition with BRAID_BIND_VALUE_UNSUPPORTED.

Runtime packages do not pull tooling, metadata, codegen, editor, or Vite dependencies. The facade also does not install database drivers; install the database driver separately. See the package map and release readiness.

The repository uses Node.js >=22.18.0 and pnpm. Run the project checks from a clean checkout before making a release decision; this README records no result for the current documentation/API revision. Publication, tagging, npm provenance, Marketplace release, and Pages deployment are separate authorized operations.