Retries

Retries are off by default. Opt in by passing retry:

const api = createClient({
  retry: { attempts: 3, backoff: "exponential" },
});

Backoff strategies

• "exponential" (default) - 100ms, 200ms, 400ms, 800ms… capped at 30 seconds
• "linear" - 100ms, 200ms, 300ms… capped at 30 seconds
• (attempt) => ms - custom function for full control

What’s retried by default

The default predicate retries transient failures only:

Error	Retried?
NetworkError	yes
5xx HTTP	yes
408 Request Timeout	yes
429 Too Many	yes
Other 4xx	no
AbortError	no
ValidationError	no

The reasoning: 4xx errors are deterministic - retrying without changing the request gives the same answer. Aborts and validation errors are explicit and never retried.

Custom retry predicate

For finer control, pass a retryOn function:

retry: {
  attempts: 5,
  backoff: "exponential",
  retryOn: (err, attempt) => {
    if (err instanceof HttpError && err.status === 503) return true;
    return false;
  },
}

The predicate gets the 1-indexed attempt number, so you can give up after a certain count even when the error type would normally retry.

Per-request override

Override the client default for one call:

await api.get("/x", {
  retry: { attempts: 0 }, // disable retries for this request
});

What you might want next

fetch-kit covers retries, response caching, and in-flight request deduplication out of the box. If you need richer server-state management - optimistic updates with rollback, normalized caches, or background refetching - that’s TanStack Query territory. fetch-kit deliberately stops short of that.