Single path
Prompts, completions, tool calls.
Model Gateway
Every prompt, completion and tool call is addressed through the Model Gateway. You choose which model class may answer which data class. There is no side door to a public endpoint.

Single path
Prompts, completions, tool calls.
Before weights
Policy sees the request first.
Class routing
Which model may see which data.
Closed
No public inference endpoint.
The application sends a prompt or tool call. Nothing bypasses the Gateway.
Policy, DLP and RBAC decide whether the call may proceed, and with which model class.
The Foundation runs the allowed model. Completions return on the same path.
The audit log writes what was asked, which model answered, and which policy fired.
Applications do not speak to weights directly. The Gateway is the address.
AgentLab actions go through the same path as a DocxIntel ask.
Vision, LLM, embeddings, reranking, assigned by policy, not by the caller.
There is no BizfyLabs-operated URL in the path, because none was built.
Pull the cable. The Gateway still mediates. Audit still writes.
DocxIntel and AgentLab share the Gateway. Two apps, one control point.
Every surface addresses the Gateway. None addresses the weights.
Allow, deny, or route to a human queue before inference.
The model that answers is one you already control.
What ran is written where it ran.
One path. No side door.
A feature that calls a public model is a perimeter hole. The Gateway is how that hole is not designed in.
DocxIntel and AgentLab do not each invent their own policy engine.
If logging is optional, it will be skipped. The Gateway makes it structural.
A gateway that requires SaaS to function is not a gateway you can air-gap.
In your office, on your documents, with the cable pulled out. No competitor's sales engineer can do this.