Concept
Engine architecture
Lumin runs in four layers. Each layer has a single responsibility, which makes the integration surface small and stable.
KP Engine
A zero-dependency Hono service in Bun. Implements VSOP87, ELP2000, Placidus houses, and 49 KP modules from scratch. No external libs beyond Hono itself. Internal-only, never exposed to the public network.
KP MCP / Connector
The public MCP gateway. Stateless per request, with a fresh server and transport per call. Handles three auth modes (anonymous, API key, OAuth), tier-based rate limiting, and response compaction. Delegates every computation to the KP Engine over the private network.
KP Agent
- The chat agent powering app.lumin.guru.
- Runs the seven-phase reading protocol via a ReAct loop.
- Compacts engine responses for LLM context.
- Streams tool results, visual blocks, and thinking events to the chat UI over SSE.
KP In-App Chat
A Next.js 16 chat UI with multi-profile support, comparison mode, and an admin panel. Talks only to the KP Agent, never directly to the engine.
Why split this way
Each layer can be deployed, scaled, and reasoned about independently. The engine is pure compute. The MCP gateway is the public auth and rate-limit surface. The agent is the reasoning loop. The frontend is the UI. A bug in one layer does not destabilize the others.