From the source
The entire stack runs locally by default.
The model, harness, conversation, and trajectory all live on the user's machine.
Work that needs the outside world, such as web search, connectors, or escalation to a stronger advisor model on the cloud, is invoked only when necessary and always gated by the user.
Sensitive data therefore never leaves the device without permission, and local models carry no inference fee: the system is private and cost-effective by construction.
An effective local-first agent requires the model and harness to be designed together.
General-purpose harnesses assume a frontier model that can absorb long contexts, navigate a broad tool surface, and plan over long horizons.
Local models are less reliable under those demands.
Rather than asking a small model to manage a harness built for a large one, we shaped the two around each other: a harness tailored to the model's capability profile, and a model post-trained to use that harness effectively.
Introduction Agentic capabilities in recent months have advanced rapidly across a wide range of knowledge-work tasks.
…