Choose inference for your agents
Agents (Beta) can use Nodus On Demand or your API key, alongside existing managed Claude routing.
Select a catalog model or named endpoint at its On Demand price, or use Anthropic and OpenAI-compatible APIs with a project Secret. Hosted inference is a coming-soon stub: the API and SDK reject it before any compute starts.
The parent and parallel subagents inherit the selected source. On Demand records actual inference receipts and shares the model budget across concurrent calls. BYOK has no extra Nodus token charge; compute remains separately billed. ModelAgent exposes the active sources in Python, and existing ClaudeAgent definitions continue to work. Consumer subscription sign-in is not part of BYOK.
See the Agents guide.
On Demand charges inference plus sandbox compute, including paid Indra/auto calls. BYOK charges only sandbox compute on Nodus; the model provider bills inference. Shared subagents use one sandbox, and recorded model calls replay without another charge.