Trust

Where is my data right now?

Source-available, self-hostable, no vendor lock-in. lynox is built so that question has a one-sentence answer.

Where your data lives

Two shapes, one principle.

Your business data sits on a host you control. What leaves it follows from how you configure lynox and what you ask it to do: the inference call to whichever LLM provider you pointed it at — the connection lasts that one request; how long the provider keeps what it received is set by its own terms — plus a web search when the agent searches, and whatever accounts or addresses you connect it to. That is not a fixed list, so the diagram below shows where data lives rather than enumerating what leaves.

Self-hosted                          Managed (lynox.ai)
─────────────                        ──────────────────
Your server                          Per-tenant container
   │                                    │
   ├── lynox process                    ├── lynox process
   └── SQLite (conversations,           ├── SQLite (same)
       knowledge graph, vault)          └── Vault: AES-256-GCM
                                            encrypted at rest

Last updated: September 4, 2026

Which LLM providers see your prompts

You pick. We never substitute.

You choose the provider at install time. The control plane never swaps one for another and never proxies through a lynox-operated relay.

Provider option	Self-hosted path	Managed path
Anthropic (Claude)	Direct from your host to `api.anthropic.com`.	Direct from your tenant container to `api.anthropic.com`. No lynox-operated proxy.
Mistral (Mistral Large / Ministral)	Direct from your host to `api.mistral.ai` (Paris, EU).	Direct from your tenant container to Mistral. No lynox-operated proxy.
Custom (OpenAI-compatible)	Whatever endpoint you point at — Ollama on localhost, LM Studio, LiteLLM, Groq, vLLM, a corporate gateway. Direct.	Same, configured per tenant.

Anthropic and Mistral commit (in their paid API terms) not to train on inference inputs. Selecting one of the optional "Efficient" or "Balanced" model strategies runs everything the engine routes by model tier on Fireworks AI (US) instead of Anthropic — chat and agent runs, and background jobs on the same tiers such as memory consolidation and knowledge-graph extraction, under a no-training obligation written into its DPA rather than its ToS — see sub-processors for what that entails. For Custom endpoints, whatever-you-point-at follows that endpoint's ToS — point at a free-tier API and you may be opting into training; point at an on-prem Ollama and the prompt stays on your machine. Sub-processor details live in the DPA.

The license keeps you free to leave

The point of Managed is convenience, not lock‑in.

lynox ships under the Elastic License 2.0 — source-available, free for any use including production and customer-facing, with one restriction: you may not resell lynox itself as a competing managed hosting service. If Managed ever stops fitting, the migration wizard exports your full instance state — vault, database, knowledge graph — to a Docker volume on a host you control. The data shape is identical either way, so there is no re-import friction.

What lynox keeps off the wire

Three things we deliberately don't do.

No third-party analytics inside lynox. The only outbound telemetry is error reporting to a first-party self-hosted Bugsink (bugs.lynox.cloud, EU, PII scrubbed): always-on for Managed under Art. 6(1)(f) legitimate interest, opt-in for self-hosted via LYNOX_BUGSINK_DSN. The marketing site uses Plausible, opt-out via cookie banner.
No automatic cloud backup. Self-hosted backups are encrypted with your own key and go where you point them (Google Drive or a local disk path today). Managed-tier backups stay inside the per-tenant Hetzner volume.
No “improve our service with your data” toggle. Inference is one-shot. Logs that retain prompt content are opt-in and disabled by default; see AI Transparency.

Dig deeper

Data sovereignty — which country processes what, sub-processor list, residency commitments.
EU compliance — GDPR mapping, lawful basis, DPA, data-subject rights.
AI transparency — EU AI Act disclosure, model list, limitations, human-oversight model.
Swiss data protection — revFADP compliance for Swiss businesses (operator is a Swiss sole proprietorship in Rapperswil-Jona).

Found a security issue?

Responsible-disclosure path: SECURITY.md on the source repo. Issues land directly with Rafael; expect a first reply within a few working days.