Key facts
| Deployment | On-prem and air-gapped with no outbound connectivity |
| Access control | Scoped API keys with rotation; enterprise SSO and RBAC options |
| Auditability | Request, model and response logging inside the boundary |
| Data grounding | Embeddings and RAG are live for doctrine, manuals and procedures |
| Models | 30+ models behind one OpenAI-compatible API |
| Change control | Model and index updates shipped as reviewed, signed artifacts on your schedule |
| Pricing model | Flat monthly self-serve plans; no per-token billing on self-serve |
| Endpoint roadmap | Audio, images, moderation, batch and fine-tuning are coming soon |
TL;DR
- Assume no outbound connectivity: air-gap models, indexes and logs.
- Control every update as a reviewed, signed artifact on your schedule.
- Scope access by classification and role, with named accountability.
- Keep humans in authority over any decision with operational effect.
- Log interactions inside the boundary for audit and investigation.
How it works, step by step
- Define the classification levels and the boundary each level requires.
- Deploy the platform air-gapped so no prompt, document or log leaves the environment.
- Establish an update process: review, sign, test and promote model and index artifacts.
- Scope keys and console access by role, with rotation and a central inventory.
- Ground answers in doctrine and manuals with citations, and refuse beyond approved content.
- Record request IDs, model versions, sources, outputs and human decisions.
- Exercise incident and failover drills inside the boundary before go-live.
Try it yourself
Open the sovereign AI readiness score →
Classification and boundary design
Start from the classification of the content, not the convenience of the tool. Each level maps to a boundary: public reference material may run on standard infrastructure, while controlled content runs inside an isolated environment with no external calls. Keep separate indexes per classification so retrieval cannot mix levels.
Document each data path and its approval. In a defense review, an undocumented path is a finding even when the destination is secure. Review the map quarterly.
Air-gapped operation and controlled updates
Air-gapped deployment runs model serving, retrieval and logging inside your network with no outbound connectivity. That creates a process question: how do improvements arrive? Treat model and index updates as artifacts that are reviewed, signed, tested against your evaluation set and promoted on your schedule. Never allow automatic updates across the boundary.
Keep a version registry so every answer can be tied to the model and index revision that produced it.
Supply chain and personnel assurance
Defense procurement scrutinizes the supply chain behind every component. Maintain an inventory of models, serving components and dependencies, with provenance and review records. Access follows the same discipline: named individuals, role-scoped keys, periodic revalidation, and immediate revocation on change. Enterprise SSO and RBAC options keep console permissions aligned with personnel systems.
Change control and human authority
Every change to a prompt, tool permission or model version goes through the same review as any operational software change. Keep evaluation evidence, known limitations and rollback plans on file. Regardless of automation level, a human commander owns any decision with operational effect; the assistant's role is retrieval, summarization and drafting from approved sources.
Honest comparison
| Control area | Plugsky capability | Common gap | Owner |
|---|---|---|---|
| Boundary | Air-gapped deployment with no outbound connectivity | External API calls from isolated networks | Security accreditation |
| Updates | Reviewed, signed artifacts promoted on your schedule | Automatic vendor updates | Configuration control |
| Access | Scoped keys, rotation, enterprise SSO and RBAC options | Shared credentials by team | Personnel security |
| Audit trail | Request, model and source logging inside the boundary | Logs stored externally | Audit |
| Grounding | RAG with citations over approved doctrine | Model answers from memory | Operations |
| Authority | Human command over operational decisions | Automated action from AI output | Command chain |
Frequently asked questions
Can the platform run without any external connectivity?
Yes. Air-gapped deployment keeps serving, retrieval and logging inside your network with no outbound calls, and the OpenAI-compatible API works locally.
How do improvements reach an air-gapped environment?
As reviewed, signed artifacts promoted through your own process. Test each update against your evaluation set before it enters production.
Is multimodal input supported?
Image and audio endpoints are coming soon. Today, process signals and scans upstream and send extracted text through the live chat and embedding APIs.
How do we handle multiple classification levels?
Run separate environments and indexes per level, with access scoped by role. Keeping levels in one shared index is the most common boundary mistake.
What does the audit trail contain?
Request IDs, model and index versions, retrieved sources, outputs and human decisions, all stored inside the boundary under your retention policy.
Do humans stay in the loop?
Yes. The assistant retrieves, summarizes and drafts. Decisions with operational effect remain with accountable humans.
How do we start?
Run a contained pilot on unclassified reference material with a single team, validate citations, logging and failover, then expand classification by classification.