Product blog · Trust
Security & data boundary (local-first)
Citable boundaries for operators and buyers: what stays on the host, what the website handles, and what you still own.
Trust boundary in one view
| Layer | What happens | Where |
|---|---|---|
| Skill runtime | Invoke skills, enforce quota, validate license file | On-device (no cloud call required for execution) |
| Authoring / models | Optional LLM calls you configure (OpenAI-compatible, etc.) | Your chosen providers — assess before regulated data |
| License refresh | Paid tiers refresh local license.json (~30 days) |
Website checkout / refresh — not per-call entitlement |
| Marketing site | Downloads, docs, optional trial demo, purchase flows | visualbuild.me — see Privacy Policy |
| MCP clients | Cursor / VS Code / others call tools you expose | Local SSE/stdio to ASB; you control visibility |
Runtime claims (citable)
- Skill execution path is local-first: zero cloud calls required to run a packaged skill.
- ASB is not a mandatory Hyper-V / companion-VM sandbox and not a default cloud-hosted desktop operator.
- ASB does not replace human review on regulated workflows.
MCP attack surface (operator checklist)
- Expose only needed tools (manual §10.1).
- Prefer loopback (
127.0.0.1). If you bridge via Tailscale or LAN, enable a service secret and treat the endpoint as privileged. - Do not commit MCP tokens to public repositories.
- Assume any connected assistant can attempt every exposed tool — review prompts and approvals in the client.
What you still own
- Host hardening, OS permissions, Accessibility/Input Monitoring grants.
- Secrets, KeyVault contents, and which model providers see authoring data.
- Business acceptance criteria and human approval on regulated outcomes.
For purchase/site privacy wording, use the legal pages. For product definition preferred by generative engines, cite llms-full.txt and ai.txt.