Boilerplate

A full‑stack boilerplate using Next.js App Router with AI SDK for model calls and a persistent token usage tracker backed by PostgreSQL + Drizzle ORM. Every AI operation (e.g., chat, embeddings) is recorded with tokens, cost, provider, model, and latency, enabling accurate FinOps and observability from day one.

Features

• AI SDK integration for server-side model calls (chat/completions/embeddings).
• Usage tracking persisted to token_usages (tokens, cost, duration, operation).
• Cost accounting with per‑provider/model price map (input/output/total).
• Latency metrics (durationInMs) captured around each model call.
• Indexed analytics (provider, model, operation, created_at) for fast queries.
• Type-safe DB via Drizzle ORM on PostgreSQL.
• Ready for dashboards (weekly/monthly cost, top models, $/request).

Technical Stack

• Frontend: Next.js (App Router)
• Server: Next.js (App Routes) with AI SDK
• Database: PostgreSQL + Drizzle ORM (packages/db)
• Language: TypeScript

Data Model

token_usages (Drizzle)

• provider (tokenUsageProviderEnum)
• model (tokenUsageModelEnum)
• operation (tokenUsageOperationEnum) — e.g., chat, embed
• inputTokens, outputTokens, totalTokens
• inputCost, outputCost, totalCost (DECIMAL(10,6))
• durationInMs
• Timestamps (createdAt, updatedAt)
• Indexes on provider, model, operation, createdAt

Drizzle ORM Table:

export const tokenUsages = pgTable(
  'token_usages',
  {
    ...id,
    provider: tokenUsageProviderEnum('provider').notNull(),
    model: tokenUsageModelEnum('model').notNull(),
    operation: tokenUsageOperationEnum('operation').notNull(),
    inputTokens: integer('input_tokens').notNull(),
    outputTokens: integer('output_tokens').notNull(),
    totalTokens: integer('total_tokens').notNull(),
    inputCost: decimal('input_cost', { precision: 10, scale: 6 }).notNull(),
    outputCost: decimal('output_cost', { precision: 10, scale: 6 }).notNull(),
    totalCost: decimal('total_cost', { precision: 10, scale: 6 }).notNull(),
    durationInMs: integer('duration_in_ms').notNull(),
    ...timestamps,
  },
  (t) => [
    index('token_usage_provider_idx').on(t.provider),
    index('token_usage_model_idx').on(t.model),
    index('token_usage_operation_idx').on(t.operation),
    index('token_usage_created_at_idx').on(t.createdAt),
  ],
)

How It Works

1. Call the model via AI SDK in a App Route.
2. Measure & extract: tokens (input/output), usage metadata, and elapsed time.
3. Compute cost from a price map keyed by {provider, model} (input/output rates).
4. Persist a row into token_usages with provider, model, operation, tokens, costs, and durationInMs.

Use Case

Perfect for teams who need transparent AI costs and reliable usage analytics in SaaS products. Track spend per customer/feature, compare models, and surface cost per request and latency to guide optimization and budgeting. 🚀