One Entity, several ways to run

Choose a foreground conversation, a scheduled runtime, or a headless worker.

An Entity is the persistent unit. A process runs it, an inference provider generates responses, and surfaces connect people and environments to it.

Mode	Command	Background daemon
Single request	uv run ngram ask rook "Hello"	No
Terminal conversation	uv run ngram talk rook	No
Full runtime	uv run ngram run rook	Yes
Headless worker	uv run ngram worker rook	Yes

The worker runs the same cognition, tools, and memory behavior as the full runtime. It omits the interactive terminal so a service can run without a TTY.

Keep one owner of active state

Connect several surfaces to the same Entity process. Do not launch duplicate workers against one file workspace or assume independent processes serialize each other’s turns. The runtime’s turn lock protects the surfaces sharing that running Entity.

An opened portable Entity also has its own canonical-state and lease rules. See portable Entities before moving a running instance between hosts.

Add presence deliberately

Start with a foreground conversation to confirm inference. Add messaging and spatial connections once that path works. Enable always-on operation when you want background schedules and proactive behavior.

The shared inference pause works across the Entity’s model consumers. It gives you a direct way to stop model use while keeping the runtime and its connections available.