Squadron · What we claim
Claims, and how to check them
Every statement here describes how this deployment is actually configured. Where something cannot be independently verified by a reader, we say so. Last reviewed: 2026-10-08.
What it is
A private AI console and API for a small, invited group. Chat, documents, web search, memory and charts
on one surface; an OpenAI-compatible /v1 API for tools. Access is by invitation.
Where the compute runs
Inference runs on accelerator hardware we own and operate on our own premises. There is no third-party model provider in the request path: the console and API talk to a local gateway, which routes only to that hardware. Embeddings for document search run on the same box.
What happens to your content
- Stored on our hardware. Chats, uploaded files and memories live in a database and file store on the machine we operate. No offsite copy exists except the encrypted backups below.
- Not written to gateway logs. The AI gateway records usage metadata (model, token counts, timestamps, key) and is configured not to log message content.
- No third-party AI subprocessors. Web search queries go to search engines through a search service we host ourselves; results are fetched by our own services.
- Visible to the operator. One person administers this deployment, and an administrator can read stored chats. That is disclosed on the privacy page and is the honest cost of a single-operator pilot.
Network path — including what we do not control
- Traffic to
chat.jmni.aiandapi.jmni.aipasses through Cloudflare, which terminates TLS at its edge and forwards to our origin over an outbound tunnel. Content therefore passes through Cloudflare's edge; we do not send it anywhere else, and we cannot make Cloudflare's handling disappear. - Only SSH is reachable on the local network; every service port is bound to loopback and published through that tunnel. The origin additionally validates Cloudflare's Access token before proxying the console.
- Network metadata (connection times, IP addresses at the edge) is necessarily visible to the edge provider and to us.
Access control
Console access is an allowlist of email addresses; sign-in is a one-time code emailed by Cloudflare, with a seven-day session per device. API access is separate: named keys with their own model list, budget and rate limits, revocable at any moment. Removing someone from the allowlist also ends their active console sessions.
Backups
Nightly backups of both databases and uploaded files are encrypted with a key that is held off this machine (in the operator's password manager) before an encrypted copy is uploaded to our own object storage. Restores are tested. Note that deleting something in the app removes it from the live system; it can persist in rotated backups until those age out.
What we do not claim
- No third-party security certification (no SOC 2 / ISO audit); no independent penetration-test report.
- No uptime SLA. This is a pilot: it is operated on a best-effort basis and may be unavailable.
- No guarantee that model output is correct; check anything that matters.
- No claim of anonymity. You are identified by the address you are invited with, and your usage is attributed to that account.
- No audit of the underlying open-source components beyond pinning versions and a weekly automated check of published security advisories against the versions we actually run.
Questions or corrections
Write to [email protected]. If a claim here turns out to be wrong, we will correct the page rather than the wording around it.