Skip to main content
Local deployment describes where the Entity runs. It can use a local Ollama model or call a hosted provider from the same computer.

​
Configure the runtime

From an application source checkout:
The local profile configures Ollama and model IDs. For a hosted model on the same machine, choose --profile hosted instead. See hosted inference. The checked-in Ollama defaults are:
These are defaults, not minimum hardware requirements. Choose a model and context budget your machine can sustain; see hardware.

​
Choose the process you need

For ask, talk, and run, --ollama starts the local Ollama service when needed. Add --pull-models only when you want to download configured models:
Once models are available, omit --pull-models for routine starts.

​
Keep state and execution explicit

Local state usually lives under ~/.ngram/entities/<Entity name>/. Use uv run ngram status rook to inspect the configured paths. Tools execute through their configured backend. Local inference does not automatically grant shell access, and a hosted model does not relocate tools to the provider. Review tool access before enabling execution on your workstation.

​
Stop cleanly

Use Ctrl+C to stop a foreground runtime. Use the shared inference pause when you want the surfaces to remain connected without model use. If you need to clean up other local ngram processes, inspect uv run ngram stop --dry-run before running it. Its scope can include the gateway and Ollama unless you pass the preservation flags described in usage and controls.