Local-first means the agent runs on hardware you control, session artifacts live in files you can grep, and your model traffic goes straight to the provider you configure — not through Vireon servers.
What leaves your machine
| Data | Destination | When |
|---|---|---|
| Prompts + tool context | Your LLM provider (e.g. OpenRouter) | Each completion |
| Optional embeddings | Embed model you configure | Recall / memory search |
| Web fetch / browser | Target URLs | When those tools run |
What stays local
- Workspace — code, edits, shell commands, git state
.agent_sessions/— append-only JSONL per task.agent_notes.json— typed memory.agent_artifacts/— distilled tool outputs- Persona —
persona/active/on disk
You can audit, back up, or delete all of it without asking us.
Approvals and destructive tools
run_shell, run_background, and similar tools use the normal approval path unless you enable YOLO in a trusted environment. Optional safety judge (fast model, single-token) can auto-clear obviously safe calls.
Air-gapped and local models
Point AGENT_API_BASE_URL at LM Studio, Ollama, or any OpenAI-compatible server. No internet required except to that endpoint.
Run it on your hardware
Install in one command. Bring your own API key. No Vireon signup.
— The Vireon Dynamics team