In the first post of this series I made the case for agreeing on an OpenAPI contract before anyone writes code. This one is about the tool that makes the contract actually pay off.

Hand-writing the frontend’s API layer is one of those jobs nobody admits to enjoying. You copy types out of the backend docs, write the same request wrapper for the fortieth time, and three weeks later someone renames a field and half your calls quietly break. Orval deletes that job. Point it at your OpenAPI spec and it generates the client code for you: typed functions for every endpoint, models that match the contract exactly, and mocks you can develop against before the backend exists.

Why the frontend team likes it

#

The mock generation is the underrated part. Once the spec is agreed, the frontend can spin up a fake server and build the entire UI against it, instead of waiting for endpoints that are perpetually “almost done”. When the real backend lands, you point the client at it and things mostly just work, because both sides were built from the same document.

The other win is that the boring code is now correct by construction. Nobody fat-fingers a field name in a hand-rolled request body, because nobody writes a hand-rolled request body anymore.

Why the backend team likes it too

#

The spec being the single source of truth changes the day-to-day conversations. Instead of chat threads asking “what does this endpoint return again?”, the frontend just reads the generated types. Backend devs get interrupted less, and they get feedback sooner: if the frontend builds against the mocks and something about the API feels awkward, you find out while it’s still cheap to change, not after it’s in production.

How it fits the workflow

#

Nothing fancy. The backend proposes the spec, both teams review it, the frontend generates the client and mocks with Orval and starts building, and the backend implements the real endpoints to the same contract. When the spec changes, both sides regenerate, and the compiler tells you exactly what broke. That last part is the whole point: drift between frontend and backend becomes a compile error instead of a production bug.

Setting it up

#

Install it as a dev dependency, along with whatever the generated code will import. For React Query hooks and mocks that’s TanStack Query, MSW and Faker:

npm install --save-dev orval msw @faker-js/faker
npm install @tanstack/react-query

Then add an orval.config.ts at the root of the project. This one is close to the example in Orval’s own docs:

import { defineConfig } from 'orval';

export default defineConfig({
  petstore: {
    input: {
      target: './openapi.yaml', // a path or a URL
    },
    output: {
      mode: 'single',
      target: './src/api/petstore.ts',
      schemas: './src/api/model',
      client: 'react-query',
      httpClient: 'fetch',
      mock: true, // MSW handlers plus Faker data
      override: {
        fetch: { includeHttpResponseReturnType: false },
      },
    },
  },
});

client decides what you get: plain fetch or axios functions, hooks for React Query, Vue Query, Svelte Query or SWR, Angular services, or Zod schemas. The includeHttpResponseReturnType: false line makes the fetch client return the response body itself rather than { data, status }, which is usually what a component wants.

Add a script so nobody has to remember the command:

{
  "scripts": {
    "api": "orval"
  }
}

npm run api writes the hooks and models. A component then uses them like any other React Query hook, with the types already filled in:

import { useListPets } from './api/petstore';

export function PetList() {
  const { data: pets, isLoading } = useListPets();
  if (isLoading) return <p>Loading…</p>;
  return (
    <ul>
      {pets?.map((pet) => <li key={pet.id}>{pet.name}</li>)}
    </ul>
  );
}

Rename a field in the spec, run npm run api again, and pet.name turns red everywhere it’s used. That’s the drift check from above, and it costs nothing to run.

Developing against the mocks

#

With mock: true, Orval also writes an MSW request handler for each endpoint, returning Faker data shaped like the response schema, plus a function that returns all of them. Hand that list to MSW’s setupWorker in the browser or setupServer in tests and the app runs against a fake backend that already matches the contract. Each handler also takes an override, so a test can return a specific payload, an empty list or an error state without anyone writing a fixture by hand.

That’s the part that lets the frontend start on day one. The mocks aren’t a separate artefact someone has to maintain; they’re regenerated from the same spec as the client, so they can’t fall behind it.

Orval vs openapi-generator, openapi-typescript and Hey API

#

They all start from an OpenAPI spec, but they stop at different places.

If you only need types, openapi-typescript is the lightest option and it’s hard to argue with. openapi-generator earns its place when the spec has to feed more than one language; its TypeScript output works, but it reads like it was written for Java first. Between Orval and Hey API it mostly comes down to taste and which one’s output you’d rather read. I keep using Orval because the mocks come out of the same run as the hooks, and that’s the feature that changed how the two teams work together.

It’s not glamorous tooling. It just quietly removes a whole category of arguments between teams, and I’ll take that trade every time.

In short, the design-first process runs once per sprint:

1. Pick two owners for the sprint’s API work, one from backend and one from frontend.
2. Together they write or update the OpenAPI schema against the sprint’s requirements, and review it until both sides agree.
3. The schema is committed to the repo. From here on it’s the contract.
4. CI stands up a mock server from the committed schema.
5. Frontend builds against the mock while backend implements the real endpoints.
6. Once the real service is deployed, the spec it generates becomes the starting point for the next sprint’s schema.

Where this came from

#

The push came from our retro. The feedback was blunt about how little visibility anyone had into API planning: backend built endpoints against their own reading of the requirements, frontend discovered the response shapes after the fact, and the gaps surfaced at integration time — the most expensive possible moment. The proposal was to pull both sides into the planning phase and make the OpenAPI schema the thing everything hangs off.

The process

#

Each week we pick two representatives, one backend and one frontend. Their job is to understand the sprint’s requirements and get the planning OpenAPI schema to match. They review the schema together, revise whatever doesn’t line up, and by the end both sides know exactly what’s being built — before anyone builds it.

Once the revisions settle, the schema gets committed to the repo, and CI/CD generates a mock API from the yaml. When the sprint wraps up, a fresh yaml is cut covering the next sprint’s API requirements, and the cycle repeats.

Tooling notes

#

For visualising the yaml, Swagger Editor does the job: upload the file and see it rendered as a proper Swagger document.

The CI/CD pipeline needs to accept a yaml and stand up a mock server from it. Prism, from Stoplight, is a widely used option and handles OpenAPI 3.1:

npm install -g @stoplight/prism-cli
prism mock openapi.yaml

That serves every path in the schema on http://127.0.0.1:4010 and answers with the examples from the schema, so GET /lists/10 returns the example list from the appendix. It also enforces the contract: because the example declares an API key, a request without an X-API-Key header gets a 401, which is exactly the behaviour the frontend should be building against. Inside a container, add -h 0.0.0.0 so it’s reachable from outside. openapi-mock is a lighter alternative that fills responses with random data.

It’s also worth linting the schema before the mock goes up, so a broken contract fails the build instead of confusing the frontend. Spectral needs a one-line ruleset to get going:

npm install -g @stoplight/spectral-cli
echo 'extends: ["spectral:oas"]' > .spectral.yaml
spectral lint openapi.yaml

Finally, this schema is a planning artefact only. Once the real API is deployed, the service should generate its own openapi.json, which you can download and feed back into the next round of planning.

Why bother

#

Because the arguments move to the cheapest possible place. With a frontend and a backend engineer in the room while the schema is written, misunderstandings get caught before they become code. The API requirements are agreed once instead of renegotiated mid-sprint, the mock server means frontend starts immediately rather than waiting, and far fewer critical details slip through — they’d have to get past two pairs of eyes and a written contract.

None of this is revolutionary. It’s just planning made concrete enough that a machine can serve it back to you as a mock — which, it turns out, is about the level of concrete a plan needs to be.

Appendix A - yaml example

#

openapi: 3.1.0
info:
  title: MyApp Sprint 3
  description: Planning schema for sprint 3. Drives the mock server only.
  version: 0.3.0
servers:
  - url: https://api-mock.example.com/v1
tags:
  - name: lists
    description: Saved lists feature (PROJ-123)
security:
  - api_key: []
paths:
  /lists:
    get:
      tags: [lists]
      summary: Return the current user's saved lists
      operationId: getLists
      responses:
        '200':
          description: OK
          content:
            application/json:
              schema:
                type: array
                items:
                  $ref: '#/components/schemas/List'
  /lists/{listId}:
    get:
      tags: [lists]
      summary: Find a list by ID
      operationId: getListById
      parameters:
        - name: listId
          in: path
          required: true
          schema:
            type: integer
            format: int64
      responses:
        '200':
          description: OK
          content:
            application/json:
              schema:
                $ref: '#/components/schemas/List'
        '404':
          description: List not found
components:
  schemas:
    List:
      type: object
      required: [id, title]
      properties:
        id:
          type: integer
          format: int64
          examples: [10]
        title:
          type: string
          examples: [Weekend reading]
        itemCount:
          type: integer
          examples: [7]
  securitySchemes:
    api_key:
      type: apiKey
      name: X-API-Key
      in: header

The waiting game

#

The traditional flow is sequential. Backend designs and builds the API, and only once the endpoints actually work does frontend start integrating. Every backend delay lands directly on the frontend’s calendar, and the frontend’s calendar is usually where the release date lives.

Contract first, code second

#

With a schema-first flow, both sides start almost at the same time. Before any code exists, the API’s structure goes into an OpenAPI schema — every endpoint, request and response model, auth method and error code. That document becomes the contract between frontend and backend.

The payoff is immediate. Everyone understands the feature the same way before it’s built. Frontend can build UI while backend is still implementing. Changes move faster because there’s one agreed source of truth to update. And you can generate a mock server straight from the schema, so frontend testing starts before a single real endpoint exists.

How it runs in practice

#

1. Gather requirements, same as for anything.
2. Design the schema: endpoints, request/response models, authentication, error codes.
3. Have both frontend and backend review it — this is where the contract earns its keep.
4. Generate a mock server (Swagger and Postman both do this) and let frontend integrate against it.
5. Backend implements the real endpoints to the same schema.
6. Swap mocks for real endpoints as they land, testing continuously so the contract holds.
7. Keep talking. The schema is a living document, not a stone tablet.

The catches

#

It isn’t free. Writing a thorough schema upfront takes real time. When requirements change, the schema has to change first, and the ripples hit both teams. And everyone needs at least a working familiarity with OpenAPI and the tooling around it.

Even so, I’ve never seen the upfront cost lose to the integration chaos it prevents. Frontend starts on day one, and the arguments happen over a document instead of over broken builds.