Run it where your data already lives.
Cloud, your own servers, a disconnected laptop on a site office, or a mix across teams. The same product in every case. What changes is where it runs and what, if anything, is allowed to leave.
Four ways to run it
We host it
The fastest way to start. We run and maintain the platform; you connect sources and work. Your data is isolated per organisation, encrypted at rest and in transit, and never used to improve anything we sell.
Fits: Teams who want to be working this week and have no residency requirement.
Your own tenant
The whole platform deployed into your cloud account, in the region you choose. Your subscription, your network, your keys. We deploy and support it; you own the boundary and the bill.
Fits: Anyone with a data-residency obligation, or a security team that wants the perimeter to be theirs.
Your own servers, no internet
Deployed into an isolated network segment with deny-by-default egress: no public address, no outbound route, private endpoints only. Your whole team reaches it over the internal network. Nothing traverses the internet.
Fits: Defence and infrastructure work, regulated clients, and anyone whose answer to "where does it go" has to be "nowhere".
One disconnected machine
The native desktop application bundles the entire stack into one installable app: the interface, the API, the background worker, the database, the federated query engine, the document indexes, the 3D model viewer, the report engine and the language layer. Everything binds to the local loopback interface. There is no server and no network dependency.
Fits: Sensitive models, site offices with no connectivity, secure rooms, and travel.
Mix them. Most organisations do.
These are not exclusive. The common shape is a hosted or on-premise platform for the office, desktop installs for site teams and travelling engineers, and an opt-in synchronisation between them that you control and can scope per user or device.
- Sync is off unless you turn it on, and both ends belong to you.
- Conflicts resolve last-writer-wins with a retained conflict log, so a disagreement is visible rather than silently discarded.
- Transport uses an authenticated key exchange with per-session keys. Secrets are re-wrapped, never sent in the clear.
- Large model files are content-addressed, so they de-duplicate across the team and their integrity is verified by hash.
- A lost laptop is cut off centrally: device-scoped tokens can be revoked without disturbing anyone else.
We do not share your data. There is no mechanism to.
This is an architectural property, not a policy we ask you to trust.
- Your data is never used to train or improve any model, ours or anyone else's.
- No usage telemetry, no analytics beacons, no crash reporting that carries your content.
- No third-party sub-processor sits in the run-time path of a self-hosted or on-premise deployment.
- Support access, where you want it, is time-boxed, audited and granted by you.
- In an isolated or disconnected deployment there is no configured outbound endpoint at all, so there is nothing to disable.
Verify it yourself
In an isolated or air-gapped deployment, "nothing leaves" is something your own network monitoring can confirm. Deny-by-default egress rules, flow logs and a packet capture will show zero outbound connections during normal operation. It is an auditable property of the deployment rather than a promise in a contract.
The language model can run inside your boundary too
This is the question most security reviews arrive at, and it has a real answer rather than a reassurance. You choose where the intelligence runs, and the choice is yours to change later.
Your own endpoint
Point the platform at the model service your organisation already runs and has already approved. Prompts and responses go where your policy already says they may go, and nowhere else.
Local models, on your hardware
Open-weight models run on a machine you own: a server with an accelerator for a team, or the workstation itself for a single analyst. Prompts, the retrieved context and the generated answers never leave the machine. No external model service is contacted, because none is configured.
A managed service
If you have no residency constraint, a hosted model service is the simplest option. Your content is not used for training, and the option remains available to move it in-house later without changing how you work.
Local models trade some capability for absolute containment: a model running on one workstation will not match the strongest hosted systems on the hardest reasoning. Which matters more is your decision, and we will tell you honestly where the line falls for your workload rather than selling you the answer that suits us.
What protects it, in every deployment
Encryption
Credentials and secret-shaped fields carry authenticated encryption at the application layer, on top of disk and storage encryption. All traffic uses TLS. On the desktop, local services bind to loopback only, so there is no network-exposed surface at all.
Keys you hold
In a self-hosted or on-premise deployment the master key lives in your key vault. We hold no copy. Rotation is supported, and key custody stays with you.
Access control
Fine-grained roles govern who can connect sources, run queries, view models, build reports and administer the platform. Organisations are isolated from one another, and licensing enforces seats and feature scope across web and desktop alike.
Single sign-on
SAML 2.0 and OIDC against your identity provider, with short-lived signed session tokens and device-scoped tokens for desktop clients, stored in the operating system keychain and revocable centrally.
Audit trail
Logins, source connections, queries, exports, report generation, administrative actions and device events are recorded with actor, action, target and timestamp.
Least privilege
Each tier runs with the minimum it needs, the network is segmented per tier, query execution is validated and constrained, and source connectivity is scoped to intended targets.
Genuine offline means never phoning home
Air-gapped operation is only real if the software never has to call out to prove it is licensed. The offline edition is activated by a cryptographically signed licence file delivered out of band, verified locally against an embedded key. There is no licence-server call, ever. Renewal is a new signed file. Disconnected deployments run indefinitely with zero outbound connectivity.
Where we stand on certification
We would rather be straight with you than imply more than we hold. We do not currently carry SOC 2 or ISO 27001 certification. What we offer instead is an architecture that removes the need to trust us with your data at all: deploy it inside your own boundary and there is no third party in the path to certify. For teams whose procurement requires certification from every vendor regardless of architecture, tell us early and we will say plainly whether we are a fit.
See the full security posture →Tell us your constraints and we will tell you what fits.
Residency, air-gap, an existing model endpoint, a mixed field-and-office team, or a security review that needs answering. Bring the hard version of the question.