Quarkus · Java 25 · Model Context Protocol

Agent intelligence.
Stateless infrastructure.

Build an OpenAI-powered supervisor that routes natural-language requests to specialist agents—and scales a fleet of MCP tool servers from zero to many without session affinity.

Run Helios locally Explore the architecture

ASKWhere is HLX-10032291?

↓

SUPERVISORRoutes to Shipment Agent

↓

MCP 01MCP 02MCP 03MCP 04MCP 05

Any request · Any replica · Same result

5domain tools

3specialist agents

No affinityrequests reach any replica

0→NKnative replicas

Choose your path

01 Run the experience Start the two services in dev mode or launch the complete five-replica topology. Get started → 02 Understand the design See how the supervisor, scoped toolboxes, load balancer, MCP fleet, and shared data fit together. View architecture → 03 Scale on OpenShift Deploy the platform, prove horizontal scaling, then switch the MCP tier to Knative scale-to-zero. Deploy Helios →

One platform, two clean responsibilities

Agent tier

The front door and the brains

The browser console reaches an OpenAI-powered supervisor. It delegates each question to a shipment, inventory, or exception specialist with only the tools that specialist needs.

• Natural-language routing
• Scoped @ToolBox bridges
• Managed MCP clients

MCP tier

Disposable compute, durable data

Identical Quarkus replicas expose validated business tools. They hold no session state, so a load balancer can send every call to a different instance.

• Streamable HTTP
• Jakarta validation
• Shared PostgreSQL state

A demo you can see, not just describe

Helios Control TowerAsk an agent, invoke raw tools, and watch successive requests rotate across ready replicas.

Ready to fly?

Launch the control tower.

Start with dev mode