AI Power Grid Provider for the Vercel AI SDK

Use AI Power Grid through the standard AI SDK interfaces for streamed text, image generation, and experimental video generation. The package also exposes a typed music-generation helper for the Grid's ACE-Step models.

Install

npm install ai @aipowergrid/ai-sdk-provider

Create a scoped API key in the Grid Console with inference.submit for generation and account.read if your application shows balances. Keep it server-side:

AIPG_API_KEY=grid_...

Do not expose the key through NEXT_PUBLIC_*, VITE_*, client components, or browser JavaScript. A server key identifies one Grid account and spends that account's credits. For a multi-user product, mint user-scoped credentials or use the Grid's delegated-user flow instead of sharing a privileged service key. Long media jobs can run for several minutes; pass timeoutMs to createAipg only when your deployment needs a tighter deadline, and use request abort signals for user cancellation.

The AIPG_API_KEY environment fallback is used only with the canonical https://api.aipowergrid.io/v1 origin. A custom baseURL is a local testing seam and requires an explicit apiKey; remote plaintext HTTP and redirects are rejected so a production credential cannot silently move to another host.

Text

import { generateText, streamText } from "ai";
import { aipg } from "@aipowergrid/ai-sdk-provider";

const result = await generateText({
  model: aipg("auto"),
  prompt: "Explain verifiable inference in two sentences.",
});

const stream = streamText({
  model: aipg("auto"),
  prompt: "Give me three names for a decentralized image app.",
});

for await (const chunk of stream.textStream) process.stdout.write(chunk);

auto lets the Grid route to an available text model. You may also pass a public text model ID returned by await aipg.listTextModels().

Images

import { generateImage } from "ai";
import { aipg } from "@aipowergrid/ai-sdk-provider";

const result = await generateImage({
  model: aipg.imageModel("Krea 2 Turbo"),
  prompt: "A solar-powered compute cooperative, editorial photography",
  size: "1024x1024",
  providerOptions: {
    aipg: { negativePrompt: "text, watermark", outputFormat: "webp" },
  },
});

Image-to-image accepts one inline AI SDK file. Remote source URLs and masks are rejected because the current Grid contract does not support them. Supported advanced aipg options are negativePrompt, steps, cfgScale, sampler, scheduler, strength, and outputFormat.

Video

AI SDK v7 marks its video interface experimental:

import { experimental_generateVideo as generateVideo } from "ai";
import { aipg } from "@aipowergrid/ai-sdk-provider";

const result = await generateVideo({
  model: aipg.videoModel("LTX Director 2.0"),
  prompt: "Slow camera push through a luminous server hall",
  resolution: "768x512",
  duration: 4,
  fps: 24,
});

Image-to-video models such as LTX-2.3 require one inline start-frame file. AI SDK frameImages, inputReferences, and remote URL inputs are rejected rather than ignored.

Music

Music generation is deliberately not presented as text-to-speech:

const song = await aipg.generateMusic({
  prompt: "Upbeat synth-rock, confident, clean harmony",
  lyrics: "Power moves across the grid",
  seconds: 30,
  bpm: 112,
  keyScale: "A minor",
  timeSignature: "4/4",
  vocalLanguage: "en",
});

Discovery and credits

const textModels = await aipg.listTextModels();
const onlineCapacity = await aipg.listOnlineModels();
const credits = await aipg.credits();
const estimate = await aipg.quote({
  model: "Krea 2 Turbo",
  modality: "image",
  n: 1,
});

Media calls preflight the live model/modality list before submitting paid work. Availability can still change between the preflight and dispatch, so callers must handle API errors.

Trust and privacy

AIPG routes jobs to remote, community-operated workers. Inputs and outputs are not end-to-end encrypted from the worker executing the job, so operators may be able to inspect plaintext content. Do not send secrets, credentials, personal data, or regulated content unless a separately documented confidential tier meets your requirements. On-chain settlement and public payout records improve economic transparency; they do not make inference private or fully trustless.

This package does not imply a partnership with Vercel. It is a community provider maintained by AI Power Grid.

Release verification

Default tests use local protocol fixtures and never spend credits. Before a release, run one bounded streamed request through the real AI SDK runtime with a disposable scoped key:

AIPG_API_KEY="..." npm run test:e2e:live

The live test records only assertions and timing through the test runner. It must never print the key, prompt, generated text, or account balance.

After that gate passes, publish only through the root publish-packages.yml workflow using a tag that exactly matches the package version, such as ai-sdk-provider-v0.1.0. Configure npm Trusted Publishing for the AIPowerGrid/grid-provider-integrations repository and that workflow; use a package-scoped NPM_TOKEN only if npm requires one for the first publication.