Skip to main content
No for hosted inference. You need enough local hardware for your selected model only when running Ollama or another local inference backend. The Entity runtime can use a hosted chat and embedding provider. See quickstart.
No. Start in a terminal or messaging app. The spatial surface also runs in a desktop browser. Headset features depend on an immersive WebXR session and the capabilities it reports.
A shell is a body and presentation layer. In the canonical ngram_entity binding, identity, memory, and reasoning stay with the running Python Entity.
Yes. The setup wizard includes hosted providers and a custom compatible endpoint route. A working chat request does not guarantee tool or embedding compatibility; validate those too. See hosted inference.
The spatial architecture separates bindings from rendering, but v1 does not ship a universal plug-and-play adapter for every external harness. The production continuity path documented here is ngram_entity. Use the protocol guide to understand the integration boundary.
No. ngram stores several forms of state and selectively retrieves, summarizes, consolidates, and decays memory. Working context is finite. Back up important state and inspect what was actually retained.
An always-on Entity can run soma, GEN, autonomous wake, distillation, consolidation, and scheduled automation work. Closing the browser does not stop its worker. See usage and controls.
No. Other background paths remain, and the daemon has a legacy initiative path when autonomy is disabled. Stop the process or service when you need a reliable end to requests.
Use the portable container and export/verify/recovery workflows. Credentials are provisioned separately. See persistence.
ngram, including the Python runtime and ngramAR, uses the MIT License. Model assets and external dependencies can have their own licenses; keep their notices when redistributing them.