ai-core-adapter-configuration

name: ai-core/adapter-configuration description: > Provider adapter selection and configuration: openaiText, anthropicText, geminiText, ollamaText, grokText, groqText, openRouterText, openaiCompatible. Per-model type safety with modelOptions, reasoning/thinking configuration, runtime adapter switching, extendAdapter() for custom models, createModel(). Generic OpenAI-compatible providers (DeepSeek, Together, Fireworks, etc.) via openaiCompatible({ baseURL, apiKey, models }) from @tanstack/ai-openai/compatible. API key env vars: OPENAI_API_KEY, ANTHROPIC_API_KEY, GOOGLE_API_KEY/GEMINI_API_KEY, XAI_API_KEY, GROQ_API_KEY, OPENROUTER_API_KEY, OLLAMA_HOST. type: sub-skill library: tanstack-ai library_version: '0.10.0' sources: - 'TanStack/ai:docs/adapters/openai.md' - 'TanStack/ai:docs/adapters/anthropic.md' - 'TanStack/ai:docs/adapters/gemini.md' - 'TanStack/ai:docs/adapters/ollama.md' - 'TanStack/ai:docs/advanced/per-model-type-safety.md' - 'TanStack/ai:docs/advanced/runtime-adapter-switching.md' - 'TanStack/ai:docs/advanced/extend-adapter.md'

Adapter Configuration

Dependency: This skill builds on ai-core. Read it first for critical rules.

Before implementing: Ask the user which provider and model they want. Then fetch the latest available models from the provider's source code (check the adapter's model metadata file, e.g. packages/ai-openai/src/model-meta.ts) or from the provider's API/docs to recommend the most current model. The model lists in this skill and its reference files may be outdated. Always verify against the source before recommending a specific model.

Setup

Create an adapter and use it with chat():

import { chat, toServerSentEventsResponse } from '@tanstack/ai'
import { openaiText } from '@tanstack/ai-openai'

const stream = chat({
  adapter: openaiText('gpt-5.2'),
  messages,
  modelOptions: {
    temperature: 0.7,
    max_output_tokens: 1000,
  },
})

return toServerSentEventsResponse(stream)

The adapter factory function takes the model name as a string literal and an optional config object (API key, base URL, etc.). The model name is passed into the factory, not into chat().

Sampling options (temperature, token limits, top_p/topP, etc.) live inside modelOptions using each provider's native key — they are not top-level options on chat(). See the per-provider table in Configuring Sampling below.

Core Patterns

1. Adapter Selection

Each provider has a dedicated package with tree-shakeable adapter factories. The text adapter is the primary one for chat/completions:

// Each factory takes model as first arg, optional config as second
import { openaiText } from '@tanstack/ai-openai'
import { anthropicText } from '@tanstack/ai-anthropic'
import { geminiText } from '@tanstack/ai-gemini'
import { grokText } from '@tanstack/ai-grok'
import { groqText } from '@tanstack/ai-groq'
import { openRouterText } from '@tanstack/ai-openrouter'
import { ollamaText } from '@tanstack/ai-ollama'

// Model string is passed to the factory, NOT to chat()
const adapter = openaiText('gpt-5.2')
const adapter2 = anthropicText('claude-sonnet-4-6')
const adapter3 = geminiText('gemini-2.5-pro')
const adapter4 = grokText('grok-4')
const adapter5 = groqText('llama-3.3-70b-versatile')
const adapter6 = openRouterText('anthropic/claude-sonnet-4')
const adapter7 = ollamaText('llama3.3')

// Optional: pass explicit API key
const adapterWithKey = openaiText('gpt-5.2', {
  apiKey: 'sk-...',
})

2. Runtime Adapter Switching

Use an adapter factory map to switch providers dynamically based on user input or configuration:

import { chat, toServerSentEventsResponse } from '@tanstack/ai'
import type { TextAdapter } from '@tanstack/ai/adapters'
import { openaiText } from '@tanstack/ai-openai'
import { anthropicText } from '@tanstack/ai-anthropic'
import { geminiText } from '@tanstack/ai-gemini'

// Define a map of provider+model to adapter factory calls
const adapters: Record<string, () => TextAdapter> = {
  'openai/gpt-5.2': () => openaiText('gpt-5.2'),
  'anthropic/claude-sonnet-4-6': () => anthropicText('claude-sonnet-4-6'),
  'gemini/gemini-2.5-pro': () => geminiText('gemini-2.5-pro'),
}

export function handleChat(providerModel: string, messages: Array<any>) {
  const createAdapter = adapters[providerModel]
  if (!createAdapter) {
    throw new Error(`Unknown provider/model: ${providerModel}`)
  }

  const stream = chat({
    adapter: createAdapter(),
    messages,
  })

  return toServerSentEventsResponse(stream)
}

3. Configuring Reasoning / Thinking

Different providers expose reasoning/thinking through their modelOptions:

import { chat } from '@tanstack/ai'
import { openaiText } from '@tanstack/ai-openai'
import { anthropicText } from '@tanstack/ai-anthropic'
import { geminiText } from '@tanstack/ai-gemini'

// OpenAI: reasoning with effort and summary
const openaiStream = chat({
  adapter: openaiText('gpt-5.2'),
  messages,
  modelOptions: {
    reasoning: {
      effort: 'high',
      summary: 'auto',
    },
  },
})

// Anthropic: extended thinking with budget_tokens
const anthropicStream = chat({
  adapter: anthropicText('claude-sonnet-4-6'),
  messages,
  modelOptions: {
    max_tokens: 16000,
    thinking: {
      type: 'enabled',
      budget_tokens: 8000, // must be >= 1024 and < max_tokens
    },
  },
})

// Anthropic: adaptive thinking (claude-sonnet-4-6 and newer)
const adaptiveStream = chat({
  adapter: anthropicText('claude-sonnet-4-6'),
  messages,
  modelOptions: {
    max_tokens: 16000,
    thinking: {
      type: 'adaptive',
    },
    effort: 'high', // 'max' | 'high' | 'medium' | 'low'
  },
})

// Gemini: thinking config with budget or level
const geminiStream = chat({
  adapter: geminiText('gemini-2.5-pro'),
  messages,
  modelOptions: {
    thinkingConfig: {
      includeThoughts: true,
      thinkingBudget: 4096,
    },
  },
})

4. Extending Adapters with Custom Models

Use extendAdapter() and createModel() to add custom or fine-tuned models while preserving type safety for the original models:

import { extendAdapter, createModel } from '@tanstack/ai'
import { openaiText } from '@tanstack/ai-openai'

// Define custom models
const customModels = [
  createModel('ft:gpt-5.2:my-org:custom-model:abc123', ['text', 'image']),
  createModel('my-local-proxy-model', ['text']),
] as const

// Create extended factory - original models still fully typed
const myOpenai = extendAdapter(openaiText, customModels)

// Use original models - full type inference preserved
const gpt5 = myOpenai('gpt-5.2')

// Use custom models - accepted by the type system
const custom = myOpenai('ft:gpt-5.2:my-org:custom-model:abc123')

// Type error: 'nonexistent-model' is not a valid model
// myOpenai('nonexistent-model')

At runtime, extendAdapter simply passes through to the original factory. The _customModels parameter is only used for type inference.

5. Configuring Sampling

Sampling controls (temperature, token limits, nucleus sampling) are passed inside modelOptions using each provider's native key. They are not top-level fields on chat()/ai()/generate().

// OpenAI — native keys
chat({
  adapter: openaiText('gpt-5.2'),
  messages,
  modelOptions: { temperature: 0.7, top_p: 0.9, max_output_tokens: 1000 },
})

// Anthropic
chat({
  adapter: anthropicText('claude-sonnet-4-6'),
  messages,
  modelOptions: { temperature: 0.7, top_p: 0.9, max_tokens: 1000 },
})

// Gemini — camelCase
chat({
  adapter: geminiText('gemini-2.5-pro'),
  messages,
  modelOptions: { temperature: 0.7, topP: 0.9, maxOutputTokens: 1000 },
})

// Ollama — NESTED under modelOptions.options
chat({
  adapter: ollamaText('llama3.3'),
  messages,
  modelOptions: {
    options: { temperature: 0.7, top_p: 0.9, num_predict: 1000 },
  },
})

Per-provider sampling keys (all live inside modelOptions):

temperature is the one key every provider names identically; token limits and some sampling options use provider-native names. Ollama nests all sampling under modelOptions.options.

6. Capability Flag: supportsCombinedToolsAndSchema

Adapters can declare an optional capability method:

supportsCombinedToolsAndSchema?(modelOptions?: TProviderOptions): boolean

When true, the engine wires outputSchema into the regular chatStream call alongside tools and harvests the schema-constrained JSON from the agent loop's final-turn text — skipping the separate structuredOutput / structuredOutputStream finalization round-trip. When false (or the method is omitted), the legacy finalization path runs.

Current per-adapter status (#605):

Subclasses can override to narrow the capability. When extending an adapter for a custom model that doesn't support the combination, return false explicitly.

6. OpenAI-Compatible Providers

Any provider that implements the OpenAI Chat Completions API (DeepSeek, Moonshot/Kimi, Together, Fireworks, Cerebras, Qwen/DashScope, Perplexity, NVIDIA NIM, LM Studio, etc.) can be used through the generic openaiCompatible factory from @tanstack/ai-openai/compatible — no dedicated package required.

import { openaiCompatible } from '@tanstack/ai-openai/compatible'
import { createModel } from '@tanstack/ai'

// Provider-factory: configure baseURL + apiKey + models ONCE,
// then select a model per call (the model arg is a type-safe union).
const deepseek = openaiCompatible({
  name: 'deepseek', // optional label for devtools/errors (default 'openai-compatible')
  baseURL: 'https://api.deepseek.com/v1',
  apiKey: process.env.DEEPSEEK_API_KEY!,
  models: [
    'deepseek-chat', // bare string → optimistic defaults: text/image in, streaming, tools, structured output
    createModel('deepseek-reasoner', {
      // rich def → precise per-model capabilities
      input: ['text'],
      features: ['reasoning', 'structured_outputs'],
    }),
  ],
})

chat({ adapter: deepseek('deepseek-chat'), messages })
chat({ adapter: deepseek('deepseek-reasoner'), messages })

config also accepts any OpenAI SDK ClientOptions (notably defaultHeaders and defaultQuery) for providers that need extra auth headers or query params.

For a single model, use the one-shot helper:

import { openaiCompatibleText } from '@tanstack/ai-openai/compatible'

chat({
  adapter: openaiCompatibleText('deepseek-chat', {
    baseURL: 'https://api.deepseek.com/v1',
    apiKey: process.env.DEEPSEEK_API_KEY!,
  }),
  messages,
})

Pass api: 'responses' to target the OpenAI Responses API instead of Chat Completions (only for the rare compatible provider that implements it, e.g. Azure OpenAI); the default is 'chat-completions', which is what nearly all compatible providers speak.

Verify the provider's current baseURL and model ids against its live docs — they drift. See docs/adapters/openai-compatible.md for the full provider table.

Common Mistakes

a. HIGH: Confusing legacy monolithic with tree-shakeable adapter

The legacy openai() (and anthropic(), etc.) monolithic adapters are deprecated. They take the model in chat(), not in the factory.

// WRONG: Legacy monolithic adapter pattern
import { openai } from '@tanstack/ai-openai'
chat({ adapter: openai(), model: 'gpt-5.2', messages })

// CORRECT: Tree-shakeable adapter, model in factory
import { openaiText } from '@tanstack/ai-openai'
chat({ adapter: openaiText('gpt-5.2'), messages })

Source: docs/migration/migration.md

b. MEDIUM: Wrong API key environment variable name

Each provider uses a specific env var name. Using the wrong one causes a runtime error:

Source: adapter source code (utils/client.ts in each adapter package).

References

Detailed per-adapter reference files:

• OpenAI Adapter
• Anthropic Adapter
• Gemini Adapter
• Ollama Adapter
• Grok Adapter
• Groq Adapter
• OpenRouter Adapter

Tension

HIGH Tension: Type safety vs. quick prototyping -- Per-model type safety requires specific model string literals. Quick prototyping wants dynamic selection with string variables. Agents optimizing for quick setup silently lose type safety. If model names come from user input or config files, use extendAdapter() to add custom names.

Cross-References

• See also: ai-core/chat-experience/SKILL.md -- Adapter choice affects chat setup
• See also: ai-core/structured-outputs/SKILL.md -- outputSchema handles provider differences transparently