@nanogpt/ai-sdk-provider

NanoGPT community provider for the Vercel AI SDK. Use the same ai package helpers (generateText, streamText, embed, generateObject, …) to talk to NanoGPT's OpenAI-compatible endpoints without rewriting application code.

Installation

npm install @nanogpt/ai-sdk-provider
# or
pnpm add @nanogpt/ai-sdk-provider

Quick start

import { generateText, embedMany } from 'ai'
import { createNanoGPT } from '@nanogpt/ai-sdk-provider'

const nanogpt = createNanoGPT({
  apiKey: process.env.NANOGPT_API_KEY!,
  // Optional overrides:
  // baseURL: 'https://nano-gpt.com/api/v1',
  // maxRetries: 5,
  // includeLegacyApiKeyHeader: false,
})

const { text } = await generateText({
  model: nanogpt.languageModel('gpt-5'),
  prompt: 'Summarise the NanoGPT product launch in two sentences.',
})

const { embeddings } = await embedMany({
  model: nanogpt.textEmbeddingModel('text-embedding-3-small'),
  values: ['NanoGPT is fast.', 'NanoGPT is affordable.'],
})

Supported capabilities

• Chat completions via languageModel() – streaming SSE, JSON mode, tool calls, seed/stop controls, and provider metadata passthrough (<NanoGPT>…</NanoGPT> envelopes).
• Text embeddings via textEmbeddingModel() – batches up to 2048 items, base64 decoding, and token usage tracking.
• Error handling consistent with the AI SDK (NanoGPTRequestError). Response metadata stays available for observability.

Image generation is not yet wired through the AI SDK interface. If you need /v1/images/generations today, call it directly with NanoGPTClient.stream/request and stay tuned for a follow-up release.

Provider options

createNanoGPT({
  apiKey: string,
  baseURL?: string,              // default: https://nano-gpt.com/api/v1
  defaultHeaders?: Record<string, string>,
  timeoutMs?: number,            // default: 60_000
  maxRetries?: number,           // default: 2
  includeLegacyApiKeyHeader?: boolean, // default: true
  fetch?: typeof fetch,
})

Per-call headers (e.g. to forward tenant IDs) can be passed from AI SDK helpers:

await generateText({
  model: nanogpt.languageModel('gpt-4.1-mini'),
  prompt,
  headers: {
    'x-nanogpt-session-id': sessionId,
  },
})

Embeddings expose provider-specific options through the standard providerOptions bag:

await embedMany({
  model: nanogpt.textEmbeddingModel('text-embedding-3-small'),
  values,
  providerOptions: {
    nanogpt: {
      encoding_format: 'base64',
      dimensions: 1024,
    },
  },
})

Scripts

• npm run build – bundle ESM, CJS, and type definitions into dist/
• npm run type-check – strict TypeScript verification
• npm run test – Vitest unit suite (chat + embeddings)
• npm run lint – ESLint for src/
• npm run format – Prettier across source, configuration, and docs

Release checklist

• Run npm run release <bump> (e.g. patch, minor, major, or an explicit version). The script checks for a clean git state, runs lint/type-check/test/build, bumps the package via npm version, publishes with npm publish --access public, and pushes commits + tags.
• For pre-release dist-tags, append npm publish flags after --: npm run release prerelease -- --tag next.

Documentation

• docs/api-reference.md – working notes mirroring NanoGPT's public API
• content/ – MDX entry for the AI SDK community provider gallery
• examples/ – sample applications showing how to swap NanoGPT into popular templates

Issues and pull requests are welcome!