> ## Documentation Index
> Fetch the complete documentation index at: https://infercrane.mintlify.site/llms.txt
> Use this file to discover all available pages before exploring further.

# TypeScript SDK

> Typed durable operations and cancellation-aware inference streams for Node.js.

The TypeScript SDK targets Node.js 20 or newer and uses the platform `fetch` implementation. It has
no runtime dependency and calls only the InferCrane HTTP surfaces.

```bash theme={"theme":{"light":"github-light-default","dark":"vesper"}}
cd sdk/typescript
npm ci
npm run build
```

<Warning>
  `@infercrane/sdk` is repository-local during release-candidate qualification. Do not depend on
  an npm package until InferCrane publishes and signs it.
</Warning>

```typescript theme={"theme":{"light":"github-light-default","dark":"vesper"}}
import { InferCrane } from '@infercrane/sdk';

const client = new InferCrane({
  apiKey: process.env.INFERCRANE_API_KEY!,
  baseUrl: 'https://infercrane.internal',
});

const operation = await client.deploy({
  name: 'qwen-prod',
  model: 'Qwen/Qwen3-8B',
  cloud: 'runpod',
  gpu: 'L40S',
  idempotencyKey: 'qwen-prod-initial',
});

await client.wait(operation.id, { timeoutMs: 900_000 });
```

`wait` accepts an `AbortSignal`. Aborting the local wait does not cancel the durable operation. Call
`client.cancel(operation.id)` only when cooperative server-side cancellation is intentional.

```typescript theme={"theme":{"light":"github-light-default","dark":"vesper"}}
const controller = new AbortController();

for await (const event of client.streamChat(
  'qwen-prod',
  [{ role: 'user', content: 'Explain this rollout.' }],
  { signal: controller.signal },
)) {
  console.log(event);
}
```

The SDK verifies HTTP status before decoding, exposes `ApiError`, `OperationFailed`,
`OperationCancelled`, `OperationTimeout`, and `StreamError`, and never retries a streaming request
after transmission begins.

## Evaluate inference evidence

```typescript theme={"theme":{"light":"github-light-default","dark":"vesper"}}
await client.setSloPolicy('qwen-prod', {
  max_ttft_p95_ms: 250,
  max_error_rate: 0.01,
});

const recommendation = await client.recommend('qwen-prod');
const history = await client.recommendations('qwen-prod', 20);
```

SLO thresholds are validated before transmission. These methods persist and inspect deterministic,
advisory decisions; they do not apply a candidate configuration automatically.