VanillaAI

Agents are becoming the way organizations use AI. VanillaAI gives them a home you control: an isolated runtime, a central gateway, and your knowledge, on your hardware.

The stack is 100% open source and runs on-premises, in private clouds, and in public clouds. It supports agentic approaches and existing open-source agents. Every layer is replaceable. Agents run isolated on a Kubernetes-based runtime on top of VanillaStack, and all model access, local or frontier, flows through one AI gateway. There are no licensing fees and no vendor lock-in.

The AI gateway is the control point of the stack. Every agent and every user reaches models through it, whether the model runs on your hardware or at a frontier provider:

• Compliance: policies decide which workload may use which model with which data
• Auditability: every request is logged and attributable
• Model routing: local models by default, frontier LLMs only when a task requires them
• Cost control: one place to see and steer what AI usage actually costs

This is what makes AI usable in regulated environments: enterprises, public sector, and government programs.

Agents run on a Kubernetes-based runtime with strict isolation per agent. The runtime supports existing open-source agents and agentic frameworks. Security comes from the architecture: separated workloads, defined network paths, no shared state between agents.

The stack runs any open-source or free-of-charge model and scales out to frontier LLMs through the gateway when a task requires it:

• Commodity infrastructure: consumer GPUs, DGX Spark, standard servers with CUDA or ROCm
• GPU sharing across workloads through Kubernetes, without expensive middleware
• Model choice per use-case, changed without re-architecting
• No token-based platform costs for local models

Agents are only as useful as the knowledge they work with. The stack connects to centralized, curated organizational knowledge that both your people and your agents consume. Structured in layers from foundational to organization-specific, kept current, without expensive model training.

This is where VanillaBrain and our AI Enablement approach plug in.

What organizations build on the stack, from running solutions to scenarios we design for enterprises and governments. Reach out to discuss your own use-case: