Backend Service Apps & API Clients

When this skill applies

Use this skill when developing a VTEX IO app that needs backend logic — REST API routes, GraphQL resolvers, event handlers, scheduled tasks, or integrations with VTEX Commerce APIs and external services.

• Building the Service entry point (node/index.ts) with typed context, clients, and state
• Creating and registering custom clients extending JanusClient or ExternalClient
• Using ctx.clients to access clients with built-in caching, retry, and metrics
• Configuring routes and middleware chains in service.json

Do not use this skill for:

• Manifest and builder configuration (use vtex-io-app-structure instead)
• GraphQL schema definitions (use vtex-io-graphql-api instead)
• React component development (use vtex-io-react-apps instead)

Decision rules

• The Service class (node/index.ts) is the entry point for every VTEX IO backend app. It receives clients, routes (with middleware chains), GraphQL resolvers, and event handlers.
• Every middleware, resolver, and event handler receives ctx with: ctx.clients (registered clients), ctx.state (mutable per-request state), ctx.vtex (auth tokens, account info), ctx.body (request/response body).
• Use JanusClient for VTEX internal APIs (base URL: https://{account}.vtexcommercestable.com.br).
• Use ExternalClient for non-VTEX APIs (any URL you specify).
• Use AppClient for routes exposed by other VTEX IO apps.
• Use MasterDataClient for Master Data v2 CRUD operations.
• Register custom clients by extending IOClients — each client is lazily instantiated on first access via this.getOrSet().
• Keep clients as thin data-access wrappers. Put business logic in middlewares or service functions.

Client hierarchy:

Architecture:

Request → VTEX IO Runtime → Service
  ├── routes → middleware chain → ctx.clients.{name}.method()
  ├── graphql → resolvers → ctx.clients.{name}.method()
  └── events → handlers → ctx.clients.{name}.method()
                               │
                               ▼
                         Client (JanusClient / ExternalClient)
                               │
                               ▼
                    External Service / VTEX API

Hard constraints

Constraint: Use @vtex/api Clients — Never Raw HTTP Libraries

All HTTP communication from a VTEX IO service app MUST go through @vtex/api clients (JanusClient, ExternalClient, AppClient, or native clients from @vtex/clients). You MUST NOT use axios, fetch, got, node-fetch, or any other raw HTTP library.

Why this matters

VTEX IO clients provide automatic authentication header injection, built-in caching (disk and memory), retry with exponential backoff, timeout management, native metrics and billing tracking, and proper error handling. Raw HTTP libraries bypass all of these. Additionally, outbound traffic from VTEX IO is firewalled — only @vtex/api clients properly route through the infrastructure.

Detection

If you see import axios from 'axios', import fetch from 'node-fetch', import got from 'got', require('node-fetch'), or any direct fetch() call in a VTEX IO service app, STOP. Replace with a proper client extending JanusClient or ExternalClient.

Correct

import type { InstanceOptions, IOContext } from '@vtex/api'
import { ExternalClient } from '@vtex/api'

export class WeatherClient extends ExternalClient {
  constructor(context: IOContext, options?: InstanceOptions) {
    super('https://api.weather.com', context, {
      ...options,
      headers: {
        'X-Api-Key': 'my-key',
        ...options?.headers,
      },
    })
  }

  public async getForecast(city: string): Promise<Forecast> {
    return this.http.get(`/v1/forecast/${city}`, {
      metric: 'weather-forecast',
    })
  }
}

Wrong

import axios from 'axios'

// This bypasses VTEX IO infrastructure entirely.
// No caching, no retries, no metrics, no auth token injection.
// Outbound requests may be blocked by the firewall.
export async function getForecast(city: string): Promise<Forecast> {
  const response = await axios.get(`https://api.weather.com/v1/forecast/${city}`, {
    headers: { 'X-Api-Key': 'my-key' },
  })
  return response.data
}

Constraint: Access Clients via ctx.clients — Never Instantiate Directly

Clients MUST always be accessed through ctx.clients.{clientName} in middlewares, resolvers, and event handlers. You MUST NOT instantiate client classes directly with new.

Why this matters

The IOClients registry manages client lifecycle, ensuring proper initialization with the current request's IOContext (account, workspace, auth tokens). Direct instantiation creates clients without authentication context, without caching configuration, and without connection to the metrics pipeline.

Detection

If you see new MyClient(...) or new ExternalClient(...) inside a middleware or resolver, STOP. The client should be registered in the Clients class and accessed via ctx.clients.

Correct

// node/clients/index.ts
import { IOClients } from '@vtex/api'
import { CatalogClient } from './catalogClient'

export class Clients extends IOClients {
  public get catalog() {
    return this.getOrSet('catalog', CatalogClient)
  }
}

// node/middlewares/getProduct.ts
export async function getProduct(ctx: Context, next: () => Promise<void>) {
  const { clients: { catalog } } = ctx
  const product = await catalog.getProductById(ctx.query.id)
  ctx.body = product
  ctx.status = 200
  await next()
}

Wrong

// node/middlewares/getProduct.ts
import { CatalogClient } from '../clients/catalogClient'

export async function getProduct(ctx: Context, next: () => Promise<void>) {
  // Direct instantiation — no auth context, no caching, no metrics
  const catalog = new CatalogClient(ctx.vtex, {})
  const product = await catalog.getProductById(ctx.query.id)
  ctx.body = product
  ctx.status = 200
  await next()
}

Constraint: Avoid Monolithic Service Apps

A single service app SHOULD NOT define more than 10 HTTP routes. If you need more, consider splitting into focused microservice apps.

Why this matters

VTEX IO apps run in containers with limited memory (max 512MB). A monolithic app with many routes increases memory usage, cold start time, and blast radius of failures. The VTEX IO platform is designed for small, focused apps that compose together.

Detection

If service.json defines more than 10 routes, warn the developer to consider splitting the app into smaller services. This is a soft limit — there may be valid exceptions.

Correct

{
  "memory": 256,
  "timeout": 30,
  "routes": {
    "get-reviews": { "path": "/_v/api/reviews", "public": false },
    "get-review": { "path": "/_v/api/reviews/:id", "public": false },
    "create-review": { "path": "/_v/api/reviews", "public": false },
    "moderate-review": { "path": "/_v/api/reviews/:id/moderate", "public": false }
  }
}

Wrong

{
  "memory": 512,
  "timeout": 60,
  "routes": {
    "route1": { "path": "/_v/api/reviews" },
    "route2": { "path": "/_v/api/reviews/:id" },
    "route3": { "path": "/_v/api/products" },
    "route4": { "path": "/_v/api/products/:id" },
    "route5": { "path": "/_v/api/orders" },
    "route6": { "path": "/_v/api/orders/:id" },
    "route7": { "path": "/_v/api/users" },
    "route8": { "path": "/_v/api/users/:id" },
    "route9": { "path": "/_v/api/categories" },
    "route10": { "path": "/_v/api/categories/:id" },
    "route11": { "path": "/_v/api/brands" },
    "route12": { "path": "/_v/api/inventory" }
  }
}

12 routes covering reviews, products, orders, users, categories, brands, and inventory — this should be 3-4 separate apps.

Preferred pattern

Define custom clients:

// node/clients/catalogClient.ts
import type { InstanceOptions, IOContext } from '@vtex/api'
import { JanusClient } from '@vtex/api'

export class CatalogClient extends JanusClient {
  constructor(context: IOContext, options?: InstanceOptions) {
    super(context, {
      ...options,
      headers: {
        VtexIdclientAutCookie: context.authToken,
        ...options?.headers,
      },
    })
  }

  public async getProduct(productId: string): Promise<Product> {
    return this.http.get(`/api/catalog/pvt/product/${productId}`, {
      metric: 'catalog-get-product',
    })
  }

  public async listSkusByProduct(productId: string): Promise<Sku[]> {
    return this.http.get(`/api/catalog_system/pvt/sku/stockkeepingunitByProductId/${productId}`, {
      metric: 'catalog-list-skus',
    })
  }
}

Register clients in IOClients:

// node/clients/index.ts
import { IOClients } from '@vtex/api'
import { CatalogClient } from './catalogClient'
import { ReviewStorageClient } from './reviewStorageClient'

export class Clients extends IOClients {
  public get catalog() {
    return this.getOrSet('catalog', CatalogClient)
  }

  public get reviewStorage() {
    return this.getOrSet('reviewStorage', ReviewStorageClient)
  }
}

Create middlewares using ctx.clients and ctx.state:

// node/middlewares/getReviews.ts
import type { ServiceContext } from '@vtex/api'
import type { Clients } from '../clients'

type Context = ServiceContext<Clients>

export async function validateParams(ctx: Context, next: () => Promise<void>) {
  const { productId } = ctx.query

  if (!productId || typeof productId !== 'string') {
    ctx.status = 400
    ctx.body = { error: 'productId query parameter is required' }
    return
  }

  ctx.state.productId = productId
  await next()
}

export async function getReviews(ctx: Context, next: () => Promise<void>) {
  const { productId } = ctx.state
  const reviews = await ctx.clients.reviewStorage.getByProduct(productId)

  ctx.status = 200
  ctx.body = reviews
  await next()
}

Wire everything in the Service entry point:

// node/index.ts
import type { ParamsContext, RecorderState } from '@vtex/api'
import { Service, method } from '@vtex/api'

import { Clients } from './clients'
import { validateParams, getReviews } from './middlewares/getReviews'
import { createReview } from './middlewares/createReview'
import { resolvers } from './resolvers'

export default new Service<Clients, RecorderState, ParamsContext>({
  clients: {
    implementation: Clients,
    options: {
      default: {
        retries: 2,
        timeout: 5000,
      },
      catalog: {
        retries: 3,
        timeout: 10000,
      },
    },
  },
  routes: {
    reviews: method({
      GET: [validateParams, getReviews],
      POST: [createReview],
    }),
  },
  graphql: {
    resolvers: {
      Query: resolvers.queries,
      Mutation: resolvers.mutations,
    },
  },
})

Common failure modes

• Using axios/fetch/got/node-fetch for HTTP calls: These libraries bypass the entire VTEX IO infrastructure — no automatic auth token injection, no caching, no retry logic, no metrics. Outbound requests may also be blocked by the firewall. Create a proper client extending ExternalClient or JanusClient instead.
• Putting business logic in clients: Clients become bloated and hard to test. Keep clients as thin wrappers around HTTP calls. Put business logic in middlewares or dedicated service functions.
• Direct client instantiation: Using new MyClient(...) inside a middleware creates clients without auth context, caching, or metrics. Always access via ctx.clients.

Review checklist

• Are all HTTP calls going through @vtex/api clients (no axios, fetch, got)?
• Are all clients accessed via ctx.clients, never instantiated with new?
• Are custom clients registered in the IOClients class?
• Does the Service entry point correctly wire clients, routes, resolvers, and events?
• Is business logic in middlewares/resolvers, not in client classes?
• Does service.json have reasonable route count (≤10)?
• Are client options (retries, timeout) configured appropriately?

Reference

• Services — Overview of VTEX IO backend service development
• Clients — Native client list and client architecture overview
• Developing Clients — Step-by-step guide for creating custom JanusClient and ExternalClient
• Using Node Clients — How to use @vtex/api and @vtex/clients in middlewares and resolvers
• Calling Commerce APIs — Tutorial for building a service app that calls VTEX Commerce APIs
• Best Practices for Avoiding Rate Limits — Why clients with caching prevent rate-limit issues