Radondocs

Providers

All 10 AI providers — env var, default model, capabilities, base URL, native() package, and quirks. OpenAI, Anthropic, Groq (free) plus seven Pro providers.

Radon AI supports 10 providers. Three are free (OpenAI, Anthropic, Groq); the other seven are Pro. Every provider implements the same interface — the differences below are absorbed inside each adapter and don't change your chat() code.

Each adapter reads its API key from apiKey in the provider config, falling back to the RADON_<PROVIDER>_API_KEY environment variable. Override any provider's base URL with the baseUrl option (for proxies, gateways, or Azure).

At a glance

SlugProviderTierWire formatEmbeddingsVision
openaiOpenAIFreeOpenAI✅✅
anthropicAnthropic (Claude)FreeAnthropic Messages—✅
groqGroqFreeOpenAI-compatible—✅
googleGoogle GeminiProGemini✅✅
mistralMistral AIProOpenAI-compatible✅✅
deepseekDeepSeekProOpenAI-compatible——
openrouterOpenRouterProOpenAI-compatible—✅
xaixAI (Grok)ProOpenAI-compatible—✅
togetherTogether AIProOpenAI-compatible✅✅
ollamaOllama (local)ProOpenAI-compatible✅✅

Capability flags are provider-can-ever

The embeddings and vision columns say whether the provider can ever do it — the specific model you pick still matters (a text-only model won't accept images). Choose an appropriate model.

OpenAI

The reference implementation of the OpenAI Chat Completions wire format. Free tier.

Env varRADON_OPENAI_API_KEY

Get a key at platform.openai.com/api-keys.

Default modelgpt-4o

Also known: gpt-4o-mini, gpt-4-turbo, o1, o1-mini.

Embeddingsyesdefault: text-embedding-3-small

Native /embeddings endpoint.

Visionyes

Image URLs and inline base64 both accepted.

Base URLhttps://api.openai.com/v1

Override via baseUrl for Azure OpenAI or a proxy.

native() packageopenai

npm i openai.

Anthropic (Claude)

The Messages API — structurally unlike OpenAI, fully bridged by Radon. Free tier.

Env varRADON_ANTHROPIC_API_KEY

Get a key at console.anthropic.com/settings/keys.

Default modelclaude-sonnet-4-6

Also known: claude-opus-4-8, claude-haiku-4-5-20251001.

Embeddingsno

No embeddings endpoint — embed() throws UnsupportedOperationError. Use a dedicated embeddings provider.

Visionyes

Image URLs and inline base64 both accepted.

Base URLhttps://api.anthropic.com/v1

Messages API version is pinned to 2023-06-01.

native() package@anthropic-ai/sdk

npm i @anthropic-ai/sdk.

Anthropic quirks

max_tokens is required — Radon supplies a default of 4096 when you omit maxTokens. There's no response_format, so structured output is best-effort via a system instruction (force a tool for a hard guarantee). stop maps to stop_sequences; frequency/presence penalties aren't supported.

Groq

An OpenAI-compatible endpoint for very fast open-model inference. Free tier.

Env varRADON_GROQ_API_KEY

Get a key at console.groq.com/keys.

Default modelllama-3.3-70b-versatile

Also known: llama-3.1-8b-instant, meta-llama/llama-4-scout-17b-16e-instruct, moonshotai/kimi-k2-instruct.

Embeddingsno

No embeddings endpoint — embed() throws UnsupportedOperationError.

Visionyes

Available on select models.

Base URLhttps://api.groq.com/openai/v1

OpenAI-compatible surface.

native() packagegroq-sdk

npm i groq-sdk.

Google Gemini

Gemini's REST shape is unlike the other two families — fully bridged. Pro tier.

Env varRADON_GOOGLE_API_KEY

Get a key at aistudio.google.com/apikey.

Default modelgemini-2.0-flash

Also known: gemini-2.0-flash-lite, gemini-1.5-pro.

Embeddingsyesdefault: text-embedding-004

Uses the batchEmbedContents endpoint.

Visionyes

Prefer inline base64 — see the URL caveat below.

Base URLhttps://generativelanguage.googleapis.com/v1beta

Auth via the x-goog-api-key header.

native() package@google/generative-ai

npm i @google/generative-ai.

Gemini quirks

Gemini cannot fetch arbitrary web image URLs — use inline base64 for vision. Tool-call ids are synthesized (call_0, call_1, …) and results are matched by tool name. Gemini accepts only a subset of JSON Schema, so Radon strips unsupported keywords from tool and structured-output schemas.

Mistral AI

OpenAI-compatible chat plus a native embeddings endpoint. Pro tier.

Env varRADON_MISTRAL_API_KEY

Get a key at console.mistral.ai/api-keys.

Default modelmistral-large-latest

Also known: mistral-small-latest, pixtral-large-latest.

Embeddingsyesdefault: mistral-embed

Native /embeddings endpoint.

Visionyes

Via Pixtral models.

Base URLhttps://api.mistral.ai/v1

OpenAI-compatible surface.

native() package@mistralai/mistralai

npm i @mistralai/mistralai.

DeepSeek

Documented drop-in OpenAI-SDK compatibility. Pro tier.

Env varRADON_DEEPSEEK_API_KEY

Get a key at platform.deepseek.com/api_keys.

Default modeldeepseek-chat

Also known: deepseek-reasoner.

Embeddingsno

No embeddings endpoint — embed() throws UnsupportedOperationError.

Visionno

No image input.

Base URLhttps://api.deepseek.com/v1

OpenAI-compatible surface.

native() packageopenai

Uses the openai SDK pointed at the DeepSeek base URL.

deepseek-reasoner returns its final answer in content; reasoning traces live in res.raw (see reasoning tokens).

OpenRouter

An OpenAI-compatible gateway to hundreds of models across vendors. Pro tier.

Env varRADON_OPENROUTER_API_KEY

Get a key at openrouter.ai/keys.

Default modelopenai/gpt-4o

Model ids are namespaced vendor/model, e.g. anthropic/claude-3.7-sonnet, google/gemini-2.0-flash-001.

Embeddingsno

The gateway routes chat, not embeddings — embed() throws.

Visionyes

Depends on the routed model.

Base URLhttps://openrouter.ai/api/v1

OpenAI-compatible surface.

native() packageopenai

Uses the openai SDK pointed at the OpenRouter base URL.

OpenRouter attribution

OpenRouter optionally attributes traffic via HTTP-Referer / X-Title headers. Set them with the referer / title provider options, or the RADON_OPENROUTER_REFERER / RADON_OPENROUTER_TITLE env vars. Always use vendor/model ids.

xAI (Grok)

OpenAI-compatible REST API for Grok. Pro tier.

Env varRADON_XAI_API_KEY

Get a key at console.x.ai.

Default modelgrok-3

Also known: grok-3-mini, grok-2-vision-1212.

Embeddingsno

No general embeddings endpoint — embed() throws.

Visionyes

Via Grok vision models.

Base URLhttps://api.x.ai/v1

OpenAI-compatible surface.

native() packageopenai

Uses the openai SDK pointed at the xAI base URL.

Together AI

Open models behind an OpenAI-compatible API, plus native embeddings. Pro tier.

Env varRADON_TOGETHER_API_KEY

Get a key at api.together.ai/settings/api-keys.

Default modelmeta-llama/Llama-3.3-70B-Instruct-Turbo

Also known: Qwen/Qwen2.5-72B-Instruct-Turbo.

Embeddingsyesdefault: BAAI/bge-base-en-v1.5

Native /embeddings endpoint.

Visionyes

Available on select models.

Base URLhttps://api.together.xyz/v1

OpenAI-compatible surface.

native() packagetogether-ai

npm i together-ai.

Ollama (local)

Runs models on your machine — a deliberate divergence from the rest. Pro tier (gating is licensing symmetry only; the adapter is built to the same bar).

API keynone

Ollama is local and needs no API key.

Default modelllama3.2

Also known: llama3.3, qwen2.5, llava (vision), nomic-embed-text. Availability depends on what you've ollama pulled.

Embeddingsyesdefault: nomic-embed-text

Native embeddings endpoint.

Visionyes

Via multimodal models like llava.

Base URLhttp://localhost:11434/v1

Set the baseUrl option (or RADON_OLLAMA_BASE_URL) for a remote host.

native() packageopenai

Uses the openai SDK pointed at the Ollama base URL (with a dummy key).

Bring your own provider

For an OpenAI-compatible endpoint, extend OpenAICompatibleProvider and register it — you get chat, streaming, tools, and embeddings for free. For anything else, implement the ModelProvider interface (or extend BaseProvider).

byo-provider.ts
import { OpenAICompatibleProvider, registerProvider } from "@radonsdk/ai";

class MyLlmProvider extends OpenAICompatibleProvider {
  readonly name = "my-llm";
  readonly defaultModel = "my-model";
  readonly capabilities = {
    chat: true, streaming: true, tools: true,
    embeddings: false, vision: false, jsonMode: true,
  };
  protected defaultBaseUrl() { return "https://llm.mycompany.internal/v1"; }
}

registerProvider("my-llm", async () => MyLlmProvider);

Custom providers are Pro-gated but built exactly like the built-ins.

Next steps

API reference

Every method, option, and return type.

Errors

Every typed error and what causes it.

Fallback chains

Automatically retry a failed chat() on backup providers — a Pro feature. How the chain works, which errors stop it, and why streaming doesn't support it.

API reference

Every method, option, and return type in @radonsdk/ai — the RadonAI class, ChatOptions, EmbedOptions, ChatResult, the StreamChunk union, and ChatStream.

On this page

At a glanceOpenAIAnthropic (Claude)GroqGoogle GeminiMistral AIDeepSeekOpenRouterxAI (Grok)Together AIOllama (local)Bring your own providerNext steps