How do I mock the OpenAI API + Anthropic API locally?

Install stunt, add a one-service block to stunt.yaml (adapter: embedded:llm-style), and run stunt up. You get a local, stateful stand-in for OpenAI API + Anthropic API OpenAI v1 / Anthropic v1 — 3 routes over rest — with no remote account, no network, and no bills.

Is the OpenAI API + Anthropic API mock data synthetic?

Yes — entirely. The adapter returns generated data only; stunt's adapter lint rejects real-looking secrets and PII at build time. It simulates the shape and behavior of OpenAI API + Anthropic API (routes, status codes, pagination, webhooks), never real records.

← all adapters

llm-style

simulates OpenAI API + Anthropic API OpenAI v1 / Anthropic v1

restadapter v0.1.03 routes

verification

SDKverified with openai-node 7.5.0 (floor)· surface 3 · not implemented 4 · differs 3

use it

# add to stunt.yaml
services:
  llm:
    adapter: embedded:llm-style

# then
$ stunt up

The embedded: source extracts this adapter from the binary — no git clone, fully offline.

how to mock the OpenAI API + Anthropic API locally

Install stunt, paste the service block above into stunt.yaml, run stunt up — then point your client at localhost instead of OpenAI API + Anthropic API. It works offline, keeps state across requests and restarts, and every response is synthetic: the route shapes, status codes, pagination, and webhook delivery follow the real OpenAI API + Anthropic API OpenAI v1 / Anthropic v1 surface — the data is fake, the behavior is real.

api surface (3)

generated from the adapter manifest — exactly what boots, nothing more · covers 0% of the real API's 246 routes (derived from sdk openai-node @ 7.5.0 (floor))

3 routes

POST/v1/chat/completions

GET/v1/models

POST/v1/messages

SDK-verified behaviors (2)

✓models.list returns the catalog
✓chat.completions.create returns a typed echo response

Not implemented (4)

No embeddings (POST /v1/embeddings) or legacy completions endpoints
No Anthropic SSE streaming (stream:true only handled for OpenAI)
No tool use, function calling, or JSON mode
No image or multimodal content blocks

Differs from the real API (3)

Responses deterministic: echo of the last user message; no model or randomness
Reply derived solely from the last user message; earlier history ignored
stream:true (OpenAI) returns one SSE chunk plus [DONE], not token-by-token

full scorecard →

want more of the OpenAI API + Anthropic API surface? adapters are YAML + Starlark — the authoring guide covers adding routes, fixtures, and state machines. Open an issue for anything big first, or just file a PR — merged routes ship in the next release.