Quickstart
This is the hero path: run an open model on your own machine and drive Claude Code against it — no GPU required.
1. Serve a model
Section titled “1. Serve a model”atlas up --model qwen3-0.6batlas up pulls the model from the starter catalog, provisions the engine runtime on first use, and
serves on http://localhost:8080. On first start Atlas mints a default API key and prints it once —
copy it from the logs.
2. Point Claude Code at it
Section titled “2. Point Claude Code at it”In another shell:
ANTHROPIC_BASE_URL=http://localhost:8080 ANTHROPIC_API_KEY=<printed key> claudeThat’s it — Claude Code now talks to your local model. The same base URL works for the Anthropic SDKs and the Claude Agent SDK; see Use with Claude Code.
One-shot, no daemon
Section titled “One-shot, no daemon”To run a single prompt without keeping a server up:
atlas run qwen3-0.6b "Write a haiku about local inference."Next steps
Section titled “Next steps”- Use with Claude Code — model aliases and default-model env vars
- Use with OpenAI SDKs & LangChain
- API compatibility — exactly what the endpoint exposes
- Deploy — move from a laptop to a GPU box or a fleet