Heimdal Nexus
A contract-first agent hub with a Game Pitch Factory module that chains local LLM and image APIs into pitch copy, mechanics, marketing text, and cover art.
Heimdal Nexus is a contract-first agent hub: one UI to see agents, tasks, and live activity, backed by a small HTTP API and optional plug-in modules. I designed it for slice-by-slice development — registry, activity API, React shell, shared inference client — so features stay isolated and removable.
The showcase module Game Pitch Factory runs a browser pipeline — SPARK → DESIGNER → SCRIBE → ARTIST → FORGE → BOXART — that chains OpenAI-compatible chat and image generation into a saved game pitch with box art, with per-step model overrides and runs that continue if you navigate away.
What I built
Architecture
I defined a layered architecture (JSON contracts, registry, Hono API, Vite/React hub, optional modules/<id>/ packages) with boundary checks so module code does not leak into the hub shell. Shared truth lives in versioned contracts; the UI reflects real task and event state from the API rather than inventing its own.
Hub UI
Built hub-ui: routing, operational/Board/Rooms home surfaces, SSE-driven activity, hub-wide inference settings, and thin AgentPage mounting for modules. Implemented the Layer B snapshot/events API and module enablement via enabled-modules.json.
Shared inference client
Created @heimdal/inference-client: a shared inference settings form (V4 capabilities: LLM + diffusion) used by the hub and modules, so local backends like LM Studio and Draw Things can be configured once and overridden per pipeline stage.
Game Pitch Factory
Delivered @heimdal/game-pitch-factory: session store, pipeline UI, prompts/parsing, save library, module-scoped CSS, and integration with hub inference baselines and per-step overlays. Optional modules register agents through manifests and ship UI from their own package — this is the most complete example.
Conventions
Documented agent-hub conventions (AGENTS.md, Cursor rules) for maintainable AI-assisted development — the project doubles as a reference for local-first AI workflows and building features in small vertical slices friendly to humans and coding agents alike.