Why Atlas
The pieces of a private inference platform exist, and other projects combine them. Atlas’s differentiation is deliberately about framing, DX, and curation rather than a novel systems idea.
What sets it apart
Section titled “What sets it apart”- Agent-first compatibility. Everyone else leads with “serve models”; Atlas leads with “point your
agent SDK at your own infra.” Change
ANTHROPIC_BASE_URLand your Claude Agent SDK / Claude Code app runs on your hardware — drop-in is a release gate, not a feature. Model aliases map SDK defaults (e.g.claude-sonnet-*) to your deployed models. - Minutes to first token. One static Go binary, one command to a working endpoint, one command to join a machine to the fleet. No Python on the host for the slim path.
- Spot-friendly GPU fleets. Workers are disposable — a spot interruption is just a heartbeat timeout, and the scheduler re-places the model on remaining capacity. Run inference at spot prices.
- Zero-inbound-ports security. GPU workers dial out, so they need no inbound rules, no public IPs, no SSH to operate. The security review takes minutes.
- Hybrid for free. Because workers dial out, the control plane can be in one place and workers anywhere — on-prem, multiple clouds, a Mac behind NAT — all behind one endpoint.
- Curated agent-capable catalog. A tested set of open models that actually do tool calling well, with correct chat templates and engine configs out of the box — where DIY setups bleed time.
- Data sovereignty. Prompts, outputs, and weights never leave hardware you control (unless you opt into cloud-fallback).
- Honest API scope. Atlas documents exactly what it doesn’t emulate (server-side tools, batches, prompt caching) rather than half-faking it — and exactly how it does support hard features like reasoning/thinking blocks. See API compatibility.
- No Kubernetes mandate. Great on bare metal and plain VMs; manifests are packaging, not architecture.
How it compares
Section titled “How it compares”- vs Ollama — Ollama-grade local DX, but for your whole fleet, with an agent-native API.
- vs DIY (vLLM + a translation layer + K8s) — the same capability without the assembly: one binary, Anthropic-native API, an agent-first catalog.
- vs cloud API routers — routers send your prompts to other people’s clouds; Atlas runs the models on yours.