Greenfield SaaS — LLM app marketplace
NexusApp — eleven AI apps you can install and chat with
The smallest possible distillation of an LLM-app-platform thesis. Each app is a chat persona with one focused job — Recipe Pal, Code Reviewer, Trip Planner, Meeting Mate. Install one from the catalog, drop into a conversation, get something useful out the other side. Replies stream token-by-token from Claude with live markdown rendering. Adding a new app is one tuple in the catalog.
Click Install on any app, then send a message. Reply streams in live.
Or open in its own tab: dev.nexu.vforce360.ai
Apps as data
An app is just a row: id, name, category, icon, accent color, starter prompt, and a system prompt. Eleven apps in four categories ship today — Everyday, Work, Creative, Learn. Adding the twelfth is one tuple in the source list; the catalog page groups it automatically.
Server-Sent streaming
Replies arrive as Anthropic content_block_delta events. The browser receives each chunk on a single SSE stream and re-renders the assistant bubble's markdown live, with a blinking caret while tokens are still flowing. No long "thinking" blob, no fake animation.
Markdown without baggage
A tiny in-page markdown renderer (paragraphs, headings, lists, fenced code, inline code, bold, italic). Zero dependencies, zero innerHTML — every text node goes through createTextNode, every structural element through createElement, so LLM output can't inject script.
How it shipped
Same three-stage deploy lane as the other examples on this page: buildah inside the worker pod, push to the in-cluster microk8s registry, helm-rotate the dev pod. The backend is ~280 lines of stdlib Python — no pip dependencies, no framework, just http.server and urllib.request piping to Anthropic. System prompts are stripped before any catalog response leaves the server, so the IP that makes each app useful stays on the inside.