LLM providers

Every major LLM. One workflow.

Frontier models from OpenAI, Anthropic, Google, Meta, Mistral, Cohere, Amazon, and xAI — plus any open-source or self-hosted model via an OpenAI-compatible endpoint. Per-role routing across providers. The connection abstraction is project-scoped and encrypted at rest. Switch models without rearchitecting.

Native integration paths

One config-table edit away from any model.

These are the integration mechanisms. Each one unlocks an entire catalog of models — Bedrock alone covers Anthropic, Meta, Mistral, Cohere, Amazon, and AI21; OpenAI-compatible reaches every open-source model in the world.

OpenAI

openai

Native web search · image input · full frontier lineup

OpenAI-compatible

openai_compatible

Connect any open-source or self-hosted model via an OpenAI-compatible endpoint

Azure OpenAI

azure

Vision deployments supported · GovCloud-eligible

AWS Bedrock

bedrock

Recommended on AWS · workload-identity-bound · frontier and open-source models in one integration · native streaming · region-aware

Google AI (Gemini)

google

Native search grounding · multimodal input

xAI (Grok)

xai

Native live search

Ollama / vLLM / llama.cpp

openai_compatible

Local and self-hosted LLMs work out of the box via OpenAI-compatible endpoints — ideal for air-gapped and on-prem deployments

Models you can use today

Every frontier model. Every open-source model. One platform.

A non-exhaustive view of what runs on L2H right now. New models are typically a same-day config update — no platform upgrade required.

OpenAI frontier models (OpenAI or Azure OpenAI)

Anthropic Claude (Bedrock or direct)

Google Gemini (Google AI)

xAI Grok (xAI)

Meta Llama (Bedrock or self-hosted)

Mistral family (Bedrock or self-hosted)

Cohere Command (Bedrock)

Amazon Nova (Bedrock)

AI21 Jamba (Bedrock)

Open-source models (DeepSeek, Qwen, Phi, Gemma, etc.) self-hosted via vLLM / Ollama

Any future model with an OpenAI- or Anthropic-compatible API

Per-role routing

The same physical credential can be wrapped in multiple connections to route per role: planner = strong model, worker = fast model, finalizer = strong again. Configure at the project level.

Planner: a strong reasoning model
Workers: a fast, parallel-friendly model
Finalizer: a strong model for consistency check
Embedded assistant: a latency-sensitive model

// Project LLM connections

{
  "connections": [
    {
      "label": "planner",
      "provider": "<your provider>",
      "model": "<strong reasoning model>",
      "role": "planner"
    },
    {
      "label": "worker",
      "provider": "<your provider>",
      "model": "<fast model>",
      "role": "worker"
    },
    {
      "label": "finalizer",
      "provider": "<your provider>",
      "model": "<strong reasoning model>",
      "role": "finalizer"
    }
  ]
}

Recommended on AWS

AWS Bedrock — workload-identity-bound (no long-lived keys), native streaming, frontier and open-source models in one place.

Recommended on Azure

Azure OpenAI — credentials managed via Azure Key Vault. GovCloud deployments supported.

Recommended for air-gap

vLLM or Ollama via OpenAI-compatible. Self-hosted models, no external network at runtime.

Have a specific model in mind?

We benchmark new model families against our 59-test KB suite as they ship. Request a tailored evaluation for your environment.

See Benchmarks Request a Demo