Gateways

Fontana provides built-in support for 23 LLM gateways, from unified routers such as OpenRouter and LLM API to direct vendor APIs, cloud-hosted foundation models on AWS Bedrock, Azure OpenAI, Microsoft Foundry, and Vertex AI, and Custom & Self-Hosted endpoints that speak the OpenAI /v1 API (Ollama, LM Studio, vLLM, llama.cpp, LocalAI, SGLang, and similar). All calls route through one gateway abstraction with Bring Your Own Key (BYOK): you supply the API keys and endpoint settings your organisation holds, and administrators approve which gateways and models are reachable in production. Embeddings run on a single active profile you select in Admin → Models → Embeddings: platform-hosted TEI or Infinity by default when those services are configured, an OpenAI-compatible gateway using your BYOK key, or Amazon Bedrock Titan using the same AWS credentials as Bedrock chat.

Azure OpenAI takes the resource name from https://your-resource.openai.azure.com and an API key. Approve a model by its deployment name. Microsoft Foundry is Azure’s multi-model catalog. Paste the OpenAI v1 endpoint (https://your-resource.services.ai.azure.com/openai/v1, or a project URL that ends in /openai/v1) and an API key, and approve a model by its deployment name. Vertex AI takes an express-mode API key for Gemini on Google Cloud. The Google gateway remains the Gemini developer API.

For retention and no-training posture on each gateway, see Zero Data Retention (ZDR).

Approved gateways

Icon	Gateway	Examples	Documentation
…	OpenRouter	600+ models via one key	openrouter.ai/docs
…	LLM API	Secondary unified gateway	llmapi.ai/docs
…	Merge Gateway	Unified routing and failover	docs.merge.dev/merge-gateway
…	Requesty	AI gateway and LLM router	docs.requesty.ai
…	OpenAI	GPT / o-series	platform.openai.com/docs
…	Anthropic	Claude	docs.anthropic.com
…	Google	Gemini	ai.google.dev
…	Groq	Fast LPU inference	console.groq.com/docs
…	Mistral	Mistral Large / Small / Pixtral	docs.mistral.ai
…	NVIDIA	Nemotron and hosted LLMs (NIM)	build.nvidia.com/docs
…	Together AI	200+ open models	docs.together.ai
…	DeepInfra	Open-source LLMs	deepinfra.com/docs
…	Fireworks AI	Serverless open models	docs.fireworks.ai
…	Baseten	Frontier open models	docs.baseten.co
…	Cerebras	Ultra-fast inference (Llama, Qwen, GPT-OSS, GLM)	inference-docs.cerebras.ai
…	xAI (Grok)	Grok chat	docs.x.ai
…	DeepSeek	DeepSeek-V3 chat and DeepSeek-R1 reasoning	api-docs.deepseek.com
…	Hugging Face	Open-weights chat via Inference Providers router	huggingface.co/docs
…	Custom & Self-Hosted	Local and self-hosted (Ollama, LM Studio, vLLM, llama.cpp, LocalAI, SGLang)	OpenAI-compatible API
…	AWS Bedrock	Claude, Llama, Titan, Nova on your AWS account	docs.aws.amazon.com/bedrock
	Azure OpenAI	OpenAI models on your Azure resource (deployment name)
	Microsoft Foundry	GPT, Claude, Llama, Mistral on one Azure endpoint
	Vertex AI	Gemini on Google Cloud (express mode)

Decision gateways

TypeSafe adds System One decision models (Jev) through its own TYPESAFE_API_KEY BYOK key. You approve Jev models in Admin → Models like any other gateway, but they answer typed questions rather than hold a conversation, so they never appear in chat model pickers and cannot be your default chat model. Assigning one to a chat agent stops the turn with a clear error.

Agents use Jev through the System One tool instead: they send the facts plus a set of multiple-choice, scale or yes/no questions (for example “which source column feeds this target field?” while building a workflow) and get an answer with a confidence for each. Turn it on or off per agent with System One under Workflow & coding on the agent card. It needs an approved Jev model and a saved TypeSafe key.

The Gateways and Models sidebar in Admin → Models lists every gateway that has a key saved, even before you approve a model or verify the key, so you can open it and add models straight away. When you choose Add model, gateways that already have models (or a key whose catalog failed to load) are listed under Configured, above the ones still waiting for a key. Jev models show the TypeSafe mark in the sidebar.