Platform · AI Assistant · Self-Hosted Models
Every AI feature in Alvor runs on a model you choose, and that model can be one you host yourself. Point the assistant at vLLM, Ollama, or any OpenAI-compatible endpoint and the agentic studios work exactly as they do with a cloud provider: the AI proposes, a person approves, and no prompt, diagram, or record ever leaves your network.
The definition
Most platforms with AI features bundle a vendor-managed model: your data goes to their AI service, on their terms, under their sub-processor agreements. Alvor never bundles a model. You connect your own provider, Anthropic, OpenAI, Google, Azure OpenAI, Amazon Bedrock, or any OpenAI-compatible endpoint, and every AI feature runs through it.
That last option is the one regulated environments care about. An OpenAI-compatible endpoint does not have to be OpenAI: it can be a vLLM server on your GPUs or an Ollama instance in your data centre, serving an open-weights model your security team has already approved. The assistant does not know or care where the endpoint lives.
Combined with on-premise deployment, this closes the loop that stalls AI adoption in defense, government, and critical infrastructure: the platform runs on your servers, the model runs on your servers, and the audit trail of every AI action stays inside the boundary. Fully air-gapped operation is supported; picture-to-diagram needs a vision-capable model, and every other studio works with any capable chat model.
How it works
vLLM or Ollama on your hardware, or any inference server exposing an OpenAI-compatible API. Your model choice, your GPUs, your network.
Model configuration is an endpoint URL and a model name, set by an administrator. Changing providers later is the same one-screen change.
Architecture diagramming, threat modeling, policy drafting, design documents, continuity planning: every studio calls the model you configured.
The assistant proposes; the write pauses on an approval card showing the exact payload; a person applies it. The model's location changes nothing about the trust model.
Every AI action is attributable and logged alongside every human action, inside your boundary.
vLLM, Ollama, and self-hosted inference servers connect the same way as the cloud providers: an endpoint and a model name.
Anthropic, OpenAI, Google, Azure OpenAI, Amazon Bedrock, or an endpoint you run. The model relationship is yours, never resold through us.
With the platform on-premise and the model inside the boundary, no AI feature needs an internet connection.
The AI proposes and drafts; a person approves every write, whatever model is behind it.
Every AI action is attributable and logged, the same as a human action, wherever the model runs.
Picture-to-diagram needs a vision-capable model; every other studio works with any capable chat model you serve.
Alvor never bundles or resells a model. The assistant sees only what the signed-in user can see, every write is approval-gated, and plan activation, incident close, and approvals stay human-only, whether the model is a cloud provider's or one on your own GPUs.
Questions
Yes. Deploy the platform on-premise and point the assistant at a model hosted inside the same boundary, over an OpenAI-compatible endpoint. In a fully air-gapped deployment nothing crosses the boundary in either direction, and every AI feature keeps working.
Get started
Whether you lead security, run IT, manage compliance, or sit in the C-suite - we'll show you your view.