shaide

Shaide is an open-source sovereign AI platform built by axem solutions that runs LLMs and AI agents entirely within your own Kubernetes infrastructure — including fully air-gapped clusters with no internet access. Licensed under Apache 2.0, the core platform is publicly available on GitHub and managed as infrastructure as code using Pulumi. It targets regulated industries, defense, public sector, and any organization that cannot send data to third-party cloud APIs.

What It Is

Shaide is a self-hosted AI platform that serves multiple AI models concurrently on Kubernetes clusters you own and control. Rather than requiring teams to assemble an inference engine, serving orchestrator, gateway, model registry, storage, and observability stack separately, shaide ships the entire stack as one installable platform. The README describes it as "distributed, multi-model LLM inference on Kubernetes you own — installed by a single command, all the way down to air-gapped clusters."

Architecture and Stack

Shaide is built in four distinct layers, each deployed as an independent Pulumi Go program:

• Platform services — An internal OCI registry (Harbor) that mirrors every container image and model weight, plus a shared Istio Gateway and Gateway API layer.
• Serving — Per-model inference stacks using vLLM as the inference engine and llm-d for multi-instance orchestration, with KV-cache-aware scheduling and inference-pool load balancing.
• Application — The shaide server providing a universal API surface, authentication, and user management, alongside a control panel UI.
• Packages — The interactive installer, observability stack, and shared Go libraries.

Deployment proceeds in a fixed order from the internal OCI registry through gateway, model serving, application layer, and optional add-ons (MCP server, monitoring). The platform exposes an OpenAI-compatible API, so any existing OpenAI-compatible SDK, agent framework, or tool works by changing only the base URL.

Sovereign and Air-Gapped Design

The platform is built specifically for organizations whose data cannot leave their infrastructure. The internal OCI registry mirrors every container image and model weight so a cluster can operate with zero egress. There is no phone-home behavior and no dependency on any vendor's cloud. The homepage states it runs "on-premise, no cloud access, no external APIs" and is designed for critical infrastructure, defense, and space sectors. The installer ships as a containerized interactive terminal that guides setup through prompts — no Helm chart archaeology or multiple READMEs to follow.

Key Capabilities

• Single-command installation onto an existing Kubernetes cluster (AWS EKS, GCP GKE, Azure AKS, on-prem RKE2)
• Distributed multi-model serving with independent scaling per model across GPU nodes
• Built-in RAG system (Knowledge Center) for secure AI retrieval from approved internal sources only
• Full logging and auditability for compliance-sensitive environments
• MCP server datasources deployed into the shared gateway
• Infrastructure as code: every layer is a versioned, reviewable Pulumi Go program
• AI Chat and Coding Agent components for engineering and operations workflows

Current Status: Early Access

The GitHub README notes that shaide is in early access as of the repository's creation in August 2026. The platform runs production workloads, but interfaces, stack configuration, and module layout may still change between releases. The project recommends pinning a specific release rather than tracking the main branch. The repository is developed across three open-source repositories: the core infrastructure and installer (shaide), the server and auth layer (shaide_server), and the web control panel (shaide_control_panel). The project is backed by Techstars according to the About page.