Agent-first compatibility
Native Anthropic Messages API plus OpenAI-compatible endpoints — the Claude Agent SDK, Claude Code, and OpenAI SDKs just work against your own hardware.
Point your agents at your own hardware. Atlas is an open source, self-hosted LLM inference
platform. It orchestrates proven inference engines (vLLM, SGLang, llama.cpp, MLX) and exposes
Anthropic-compatible (/v1/messages) and OpenAI-compatible APIs, so existing tooling works with a
one-line config change: ANTHROPIC_BASE_URL=https://your-atlas.
View on GitHub
Agent-first compatibility
Native Anthropic Messages API plus OpenAI-compatible endpoints — the Claude Agent SDK, Claude Code, and OpenAI SDKs just work against your own hardware.
Your hardware, anywhere
A lightweight worker runs on each machine and dials out to the control plane. No inbound ports, no Kubernetes required.
Ollama-grade DX, fleet-grade scale
One binary: atlas up, pull a model, serve it. The same binary scales out to a multi-node GPU
cluster.
Best engine for the job
Atlas orchestrates vLLM, SGLang, llama.cpp, and MLX rather than reinventing them.
Get started
Use it
Claude Code, OpenAI SDKs & LangChain, and the model catalog.
Deploy
Laptop, Docker, a GPU box, or a fleet — one binary, one endpoint.
Operate
Keys, observability, and TLS.