Skip to Oxygen content

Local-First AI: What Actually Leaves Your Machine

“Local AI” is meaningless if you cannot say what stays and what leaves. Oxygen's split is deliberate: files and databases stay on your machine — OS-sandboxed in the native build, or in Docker containers in the Docker build. The main thing that leaves is model context for the LLM calls you choose to make.

What stays local

  • Organization corpus uploads in on-device storage (native data directory or Docker volumes).
  • Graph and vector stores used for document/CSV intelligence.
  • Sandbox environment, packages, and workspace mounts.
  • Editors, scripts, and in-app browser session clutter (no history/bookmarks theater).

What can leave

When you use a hosted model, prompts and the context assembled for that turn go to the model gateway you enable. That is the intentional egress path — metering and BYOK exist so finance/IT can reason about it. It is not the same as pasting a client folder into a consumer chatbot with training ambiguity.

Why isolation matters

The native build runs the data plane and sandbox as OS-sandboxed processes (macOS Seatbelt, Linux systemd, Windows AppContainer) so writes stay in the app data directory and any workspace you mount. The Docker build does the same job with containers and volumes. Either way, org uploads live under your disk policies — closer to “workstation software” than “mystery cloud bucket.”

Practical habits

  1. Prefer corpus/workspace workflows over copy-paste into random web UIs.
  2. Use BYOK on paid plans when keys must stay under your control.
  3. Minimize PII in analytics uploads (especially HR).
  4. Verify numbers and client-facing language before send.

FAQ

Do my files get trained into a public model?
Oxygen keeps corpus files on your machine. Hosted LLM providers’ training policies depend on the gateway you enable — prefer BYOK and enterprise terms when that matters.
Can I work fully offline?
Local storage and sandbox stay on-device. Hosted model calls need network. Exact offline model options depend on how you configure providers.