Skip to main content
Use a worker for an Entity that should remain available as a service:
The CLI requires the Entity argument. In the supplied Railway configuration, NGRAM_ENTITY supplies that argument to the service start command.

What the worker owns

The worker loads the Entity declaration and state, connects configured messaging platforms, starts the presence daemon, and makes the Entity bridge available when configured. It does not launch an interactive terminal conversation. Inference can be local, private-gateway, or hosted. Tools use their configured execution backend. Moving inference to an API provider does not move shell execution or memory ownership there.

Before you leave it running

  • Verify both chat and embeddings.
  • Confirm the actual database and file paths are durable.
  • Restrict messaging access and configure an operator account.
  • Check the inference pause and resume controls.
  • Review autonomous wake limits and scheduled routines.
Use the Railway deployment guide for the supplied entrypoint and service roles, or supervise the same worker command with your own process manager.

Observe background work

Worker logs include autonomous wake start, round completion, and session completion events. The optional detailed banner includes body and wake context. Treat logs and autonomy transcripts as potentially private. autonomy.wake_verbose_worker_log controls the large worker banner. autonomy.wake_user_visible_status, tool-activity settings, and Telegram /wakequiet affect presentation. They do not disable inference.

Pause, restart, or stop

Use spatial Pause inference or Telegram /pause to block model requests across the Entity. The marker persists beside the journal, so a correctly configured persistent worker remains paused after a restart. To stop the process completely, use the service manager on the host running it. Stopping a separate ngram api health service or closing a browser does not stop the worker. After changing private environment variables or startup configuration, restart through that service manager and verify the active provider and state paths again.