Providers

Amodal supports multiple LLM providers with a unified interface, built on the Vercel AI SDK. Switch providers by changing a field in amodal.json — no code changes needed.

Supported Providers

Provider	Models	Auth env var	Notes
Anthropic	Claude Opus 4, Sonnet 4, Haiku 4.5	`ANTHROPIC_API_KEY`	Native adapter, prompt caching supported
OpenAI	GPT-4o, GPT-4, o1, o3	`OPENAI_API_KEY`	Native adapter
Google	Gemini 3.1 Pro, 3 Flash, 2.5 Pro/Flash	`GOOGLE_API_KEY`	Native adapter, same key as webTools
Groq	Llama 3.3 70B, Kimi-K2	`GROQ_API_KEY`	OpenAI-compatible — ultra-low latency
DeepSeek	deepseek-chat, deepseek-reasoner	`DEEPSEEK_API_KEY`	OpenAI-compatible
xAI	Grok-4, Grok-3	`XAI_API_KEY`	OpenAI-compatible
Mistral	Mistral Large, Small	`MISTRAL_API_KEY`	OpenAI-compatible
Fireworks	Open-weight models (Llama, Qwen, etc.)	`FIREWORKS_API_KEY`	OpenAI-compatible hosting
Together	Open-weight models	`TOGETHER_API_KEY`	OpenAI-compatible hosting
Bedrock	Claude, Titan, Llama (AWS)	AWS credentials	Available for eval judges; agent loop coming
Azure OpenAI	GPT-4, GPT-4o on Azure	Azure creds	Available for eval judges; agent loop coming
Custom	Any OpenAI-compatible endpoint	per-provider	Set `baseUrl` to point at your own inference server

Recommended models by use case

Use case	Model	Input/Output per 1M	Why
Production agents	`gemini-3.1-pro-preview`	$2.00 / $12.00	Best quality/cost ratio, strong tool-calling
Budget agents	`gemini-3-flash-preview`	$0.50 / $3.00	Fast, cheap, good for simple workflows
Maximum quality	`claude-opus-4-20250514`	$15.00 / $75.00	Best reasoning, prompt caching cuts repeat costs
Fast + cheap	`gemini-3.1-flash-lite-preview`	$0.25 / $1.50	Sub-agent dispatch, simple tasks
Low-latency	`groq/llama-3.3-70b-versatile`	~$0.60 / $0.80	Groq hardware, ~300 tok/s

OpenAI-compatible providers (Groq, DeepSeek, xAI, Mistral, Fireworks, Together) reuse the OpenAI adapter with a per-provider baseUrl. Any additional OpenAI-compatible endpoint works by setting baseUrl explicitly in amodal.json — no code changes needed.

Configuration

Auto-detection

Set the relevant environment variable and Amodal auto-detects the provider:

export ANTHROPIC_API_KEY=sk-ant-...
amodal dev   # uses Anthropic automatically

Explicit config

Specify under models.main in amodal.json:

{
  "models": {
    "main": {
      "provider": "anthropic",
      "model": "claude-sonnet-4-20250514"
    }
  }
}

For OpenAI-compatible providers, set provider to the provider name and optionally override baseUrl:

{
  "models": {
    "main": {
      "provider": "groq",
      "model": "llama-3.3-70b-versatile",
      "apiKey": "env:GROQ_API_KEY"
    }
  }
}

Failover

Every ModelConfig supports a fallback chain — another ModelConfig tried when the primary fails (5xx, 429, timeout, auth error):

{
  "models": {
    "main": {
      "provider": "anthropic",
      "model": "claude-sonnet-4-20250514",
      "fallback": {
        "provider": "openai",
        "model": "gpt-4o"
      }
    }
  }
}

Failover is per-request. If Anthropic recovers, the next request goes back to the primary. For programmatic use, the runtime also exports createFailoverProvider() with configurable retry count and backoff.

Streaming

All providers support streaming via the LLMProvider.streamText() interface. The streaming shape is unified — state handlers don't know which provider is active.

Multi-Model Comparison

Use amodal eval or amodal ops experiment to compare providers:

amodal eval --providers anthropic,openai,google

This runs the same eval suite against each provider and reports quality, latency, and cost differences.