AI Wrapper on Next.js 16 (App Router) + Postgres (Neon) + Better Auth

✓ Verified 2026-06-21 · next 16.2.9 · postgres 3.4.9 · better-auth 1.6.20

Type-checked against the real SDKs, migration applied to a live Postgres, pooling tested — then tracked for upstream drift and re-verified when it moves. How we verify →

What you're getting

Setup

bun add next react react-dom drizzle-orm postgres better-auth

Environment variables:

Apply the schema: bunx drizzle-kit push

Initialization

src/lib/db.ts — Database client

import { drizzle } from "drizzle-orm/postgres-js";
import postgres from "postgres";

// Neon pooled endpoint = PgBouncer transaction mode → prepared statements off.
// ponytail: single module-level client; the serverless runtime + PgBouncer do
// the pooling, so no custom pool/globalThis singleton dance needed.
const client = postgres(process.env.DATABASE_URL!, { prepare: false });

export const db = drizzle({ client });

src/lib/auth.ts — Auth instance

// Better Auth instance (self-hosted, Next.js 16 (App Router)).
import { betterAuth } from "better-auth";
import { drizzleAdapter } from "better-auth/adapters/drizzle";
// Reuse the SAME postgres-js/Drizzle client the db slice exported in
// src/lib/db.ts — Better Auth shares the pooled `DATABASE_URL` connection.
import { db } from "./db";

export const auth = betterAuth({
  // Drizzle adapter over the shared client; provider "pg" => Postgres DDL
  // for Better Auth's own user/session/account/verification tables.
  database: drizzleAdapter(db, { provider: "pg" }),
  // ponytail: email+password is the shortest real auth that works out of
  // the box — add socialProviders / plugins here when the app needs them.
  emailAndPassword: { enabled: true },
  secret: process.env.BETTER_AUTH_SECRET,
  baseURL: process.env.BETTER_AUTH_URL,
});

export type Session = typeof auth.$Infer.Session;

app/api/auth/[...all]/route.ts — Route handler

// Mount Better Auth on Next's route layer.
import { toNextJsHandler } from "better-auth/next-js";
import { auth } from "@/lib/auth";

export const { GET, POST } = toNextJsHandler(auth);

proxy.ts — Route protection

// Session-protection proxy for Next.js 16 (App Router) (Next 16 renamed middleware.ts → proxy.ts).
// ponytail: getSessionCookie only checks the cookie EXISTS (no DB hit at
// the Edge runtime) — do the real auth.api.getSession() check inside
// protected Server Components / route handlers. This just bounces
// logged-out users before render.
import { NextResponse, type NextRequest } from "next/server";
import { getSessionCookie } from "better-auth/cookies";

export function proxy(request: NextRequest) {
  const sessionCookie = getSessionCookie(request);
  if (!sessionCookie) {
    return NextResponse.redirect(new URL("/sign-in", request.url));
  }
  return NextResponse.next();
}

export const config = {
  // ponytail: guard the SaaS app surface; widen the matcher per app-type.
  matcher: ["/dashboard/:path*", "/settings/:path*"],
};

AI app schema: user profiles, token metering & prompt logs

import { relations, sql } from "drizzle-orm";
import {
  bigint,
  check,
  index,
  integer,
  jsonb,
  pgTable,
  text,
  timestamp,
  uuid,
} from "drizzle-orm/pg-core";
// Better Auth owns identity; we only reference its `user` table by id.
import { user } from "./auth-schema";

export type CallStatus = "ok" | "error" | "filtered";

/** One profile per Better Auth user — extends identity with AI-app preferences
 *  and the soft monthly token budget the metering layer enforces. */
export const userProfiles = pgTable(
  "user_profiles",
  {
    id: uuid("id").primaryKey().defaultRandom(),
    // Better Auth's user.id is text — match it, don't recast. One profile per
    // user, so the FK itself is unique (this is the 1:1 extension of `user`).
    userId: text("user_id")
      .notNull()
      .unique()
      .references(() => user.id, { onDelete: "cascade" }),
    displayName: text("display_name"),
    // Default model this user's requests fall back to (e.g. gpt-4o-mini).
    defaultModel: text("default_model").notNull().default("gpt-4o-mini"),
    // Soft cap the metering layer checks token_usage against per calendar month.
    monthlyTokenBudget: bigint("monthly_token_budget", { mode: "number" })
      .notNull()
      .default(0),
    createdAt: timestamp("created_at", { withTimezone: true })
      .notNull()
      .defaultNow(),
  },
  (t) => [index("idx_profile_model").on(t.defaultModel)],
);

/** Append-only token meter: one row per upstream LLM call. Roll up by
 *  user + window for budget enforcement and cost reporting. */
export const tokenUsage = pgTable(
  "token_usage",
  {
    id: uuid("id").primaryKey().defaultRandom(),
    userId: text("user_id")
      .notNull()
      .references(() => user.id, { onDelete: "cascade" }),
    // Which upstream model was billed (provider model id, opaque to us).
    model: text("model").notNull(),
    promptTokens: integer("prompt_tokens").notNull().default(0),
    completionTokens: integer("completion_tokens").notNull().default(0),
    // ponytail: store cost in micro-cents as an integer — no float money, and
    // fine-grained enough for sub-cent per-token pricing without a numeric type.
    costMicrocents: bigint("cost_microcents", { mode: "number" })
      .notNull()
      .default(0),
    createdAt: timestamp("created_at", { withTimezone: true })
      .notNull()
      .defaultNow(),
  },
  (t) => [
    // Drives the "tokens used this month" rollup + budget check.
    index("idx_usage_user_time").on(t.userId, t.createdAt),
    index("idx_usage_model").on(t.model),
  ],
);

/** Append-only prompt/completion history — the auditable record of every
 *  exchange (prompt in, completion out, model, token split, outcome). */
export const promptLogs = pgTable(
  "prompt_logs",
  {
    id: uuid("id").primaryKey().defaultRandom(),
    userId: text("user_id")
      .notNull()
      .references(() => user.id, { onDelete: "cascade" }),
    model: text("model").notNull(),
    prompt: text("prompt").notNull(),
    // Null while a streamed completion is still in flight.
    completion: text("completion"),
    promptTokens: integer("prompt_tokens").notNull().default(0),
    completionTokens: integer("completion_tokens").notNull().default(0),
    latencyMs: integer("latency_ms"),
    status: text("status").$type<CallStatus>().notNull().default("ok"),
    // ponytail: opaque per-call metadata (tool calls, finish_reason, etc.) —
    // jsonb bag beats a column-per-provider-field churn.
    metadata: jsonb("metadata"),
    createdAt: timestamp("created_at", { withTimezone: true })
      .notNull()
      .defaultNow(),
  },
  (t) => [
    // Drives the per-user history feed (most-recent-first).
    index("idx_log_user_time").on(t.userId, t.createdAt),
    check(
      "prompt_logs_status_check",
      sql`${t.status} in ('ok','error','filtered')`,
    ),
  ],
);

export const userProfilesRelations = relations(userProfiles, ({ one }) => ({
  user: one(user, {
    fields: [userProfiles.userId],
    references: [user.id],
  }),
}));

export const tokenUsageRelations = relations(tokenUsage, ({ one }) => ({
  user: one(user, { fields: [tokenUsage.userId], references: [user.id] }),
}));

export const promptLogsRelations = relations(promptLogs, ({ one }) => ({
  user: one(user, { fields: [promptLogs.userId], references: [user.id] }),
}));

Connection & security

## Next.js 16 ↔ Postgres pooling (Neon pooled endpoint, PgBouncer transaction mode)

Connect through Neon's **pooled** endpoint (`-pooler` host) via `DATABASE_URL`. Serverless
functions are short-lived and concurrent, so PgBouncer in **transaction mode** is what keeps
Postgres' connection ceiling from being blown.

### `prepare: false` is mandatory
Transaction-mode PgBouncer hands each transaction a different backend, so server-side prepared
statements (postgres-js' default) silently break across the pool. Disable them on the client:
`postgres(url, { prepare: false })`. This is also why **Drizzle, not Prisma**, is paired here —
Prisma's prepared-statement reliance is a blocked intersection on this endpoint.

### Connection reuse
- Construct the postgres-js client at **module scope** (`src/lib/db.ts`) so warm function
  instances reuse one socket instead of opening one per request.
- Cap the driver pool small — `max: 1` per instance. The shared pool lives in PgBouncer, not in
  your function; a large per-instance `max` just multiplies idle connections across instances.
- Keep `idle_timeout` ~20s and `connect_timeout` ~10s so frozen instances release backends fast.

### No session-level features
Transaction mode forbids anything that spans transactions on one backend: `LISTEN/NOTIFY`,
session-scoped `SET`, advisory-lock sessions, server-side cursors, and `WITH HOLD`. Need any of
those? Use Neon's **direct** (non-pooled) endpoint for that path only.

### Thresholds
- Drizzle/postgres-js: `prepare: false`, `max: 1`, `idle_timeout: 20`, `connect_timeout: 10`.
- Neon Free pooled budget is ~10k client connections; keep concurrency well under the project's
  `max_connections` (often 100–900 by plan) by leaning on PgBouncer, never on driver pooling.

Decisions & compatibility

Related stacks