@apimaster/ai-sdk-provider

Vercel AI SDK provider for APIMaster and any other OpenAI-compatible gateway — plus generateVideo, which the AI SDK has no model type for.

npm i @apimaster/ai-sdk-provider ai zod

0.2.x targets AI SDK 7 and needs Node 22 or later. On AI SDK 5, use @apimaster/[email protected].

import { apimaster } from '@apimaster/ai-sdk-provider';
import { generateText } from 'ai';

const { text } = await generateText({
  model: apimaster('gpt-5.5'),
  prompt: 'Why does time to first token matter?',
});

Set APIMASTER_API_KEY in the environment, or pass a key explicitly:

import { createAPIMaster } from '@apimaster/ai-sdk-provider';

const apimaster = createAPIMaster({
  apiKey: process.env.MY_KEY,
  baseURL: 'https://apimaster.ai/v1',   // any OpenAI-compatible gateway
});

What you get

| Call | Backed by | | --- | --- | | apimaster(id) / apimaster.chatModel(id) | @ai-sdk/openai-compatible chat model | | apimaster.completionModel(id) | completion model | | apimaster.embeddingModel(id) | embedding model | | apimaster.imageModel(id) | image model (generateImage from ai) | | apimaster.generateVideo({...}) | submit + poll + MP4 URL |

Chat, completion, embedding and image models come straight from @ai-sdk/openai-compatible. This package deliberately does not reimplement them — it adds the preset base URL, the key lookup, model-id autocomplete, and video.

Streaming, tools, structured output

Everything the AI SDK does works unchanged:

import { streamText, generateObject } from 'ai';
import { z } from 'zod';

const stream = streamText({ model: apimaster('claude-sonnet-4-6'), prompt: 'Count to ten' });
for await (const chunk of stream.textStream) process.stdout.write(chunk);

const { object } = await generateObject({
  model: apimaster('gpt-5.5'),
  schema: z.object({ name: z.string(), age: z.number() }),
  prompt: 'Invent a character',
});

One setting worth getting right

Most models on this gateway are reasoning models: they spend part of the token budget on hidden reasoning before emitting visible text. A tight maxOutputTokens returns an empty string, not a short answer. Measured on one model, a 64-token budget was consumed 59 tokens by reasoning; at 256 it answered correctly. If output is empty, raise the budget before assuming the model is broken.

Images

import { generateImage } from 'ai';

const { image } = await generateImage({
  model: apimaster.imageModel('gpt-image-2'),
  prompt: 'a corgi astronaut on the moon',
  size: '1792x1024',
});

Image generation is slow — a 1K render measured 160 s, and 4K can take ten minutes. Give the request a generous timeout.

Video

const video = await apimaster.generateVideo({
  model: 'seedance-2.5',
  prompt: 'a waterfall forming a rainbow, cinematic',
  durationSeconds: 4,
  aspectRatio: '16:9',
});

console.log(video.url, video.elapsedMs);

Video is slow: seedance-2.5 took about 15 minutes for a 4-second clip, so generateVideo waits up to 30 minutes by default (maxWaitMs).

Image-to-video:

const video = await apimaster.generateVideo({
  prompt: 'slow push-in, hair moving in the breeze',
  referenceImageUrl: 'https://example.com/face.jpg',
  aspectRatio: '9:16',      // always explicit: a portrait reference with no aspect
  download: true,           // ratio is treated as 16:9 and comes back letterboxed
});

await fs.writeFile('clip.mp4', video.bytes!);

The returned URL needs the API key as a bearer token and expires, so download: true is the safer option when you intend to keep the file.

If the job outlives maxWaitMs, the error includes the task id — the job is not lost and you can keep polling GET /v1/videos/{id} yourself.

Model ids

The ids in the types are the ones verified on the default gateway; any other string is accepted, so a model added after this release still works. List what an endpoint serves:

curl -s https://apimaster.ai/v1/models \
  -H "Authorization: Bearer $APIMASTER_API_KEY" | jq -r '.data[].id'

Development

npm install
npm run typecheck
npm run build && npm test

Tests inject a stub fetch, so they need no key, no network and spend nothing.

License

MIT