Skip to content

Architecture

prompy/
├── web/ Next.js 16 app (App Router, Server Actions, Tailwind v4)
├── docs/ This site — Astro + Starlight
└── supabase/ SQL migrations (schema, row-level security, RPC functions)
Concern Choice
App Next.js on Vercel (Fluid Compute, Node.js runtime)
Database & auth Supabase Postgres + Supabase Auth (email + Google) — free tier
Images imgbb API, called server-side from /api/upload
MCP mcp-handler at /api/mcp, Streamable HTTP
AI news /api/ai-news, aggregated from public feeds and cached for 15 minutes
Theming next-themes, light / dark / system
Docs Astro Starlight
  • profiles — one per user, created by a trigger on sign-up.
  • prompts — owner_id, optional workspace_id, tags text[], visibility.
  • workspaces, workspace_members — roles owner, editor, viewer.
  • workflows, workflow_steps — multi-step workflow templates (migration 0018). A workflow is shared like a prompt (visibility, optional workspace_id; can_read_workflow mirrors can_read_prompt). Steps follow their workflow: workspace owners and editors edit, viewers read. save_workflow_steps(wid, steps) is a security invoker function that replaces a workflow’s steps in one transaction. Admins (is_admin()) can read every workflow and its steps, unpublish public ones and delete any (migration 0019).
  • learning_resources — the Learning Hub at /courses (migration 0020). Everyone, signed out included, reads rows with is_published; admins (is_admin()) read drafts and are the only ones who can insert, update or delete. A trigger normalises topics (lowercase, unique, at most 8), keeps youtube_id only on videos and fixes created_by. url is unique.
  • design_templates — Design Templates, curated at /admin/templates (migration 0024). Same read rule as learning_resources: published rows for everyone, drafts for admins, writes for admins only. Column grants keep copy_count out of reach of the app; only increment_template_copy(id) (security definer, callable by anyone, published rows only) adds one. A trigger normalises style tags (lowercase, unique, at most 6) and the palette (lowercase #rrggbb, at most 8), checks that screenshots only has desktop, mobile and dark keys with https or /templates/ links, fixes created_by, and leaves updated_at alone when only the copy count changes. slug is unique.
  • likes, comments — likes and new comments are limited to public prompts; comment reads also require access to the referenced prompt.
  • api_tokens — SHA-256 hashes of MCP tokens.
  • activity_logs — append-only admin, prompt and authentication events; readable only by admins.
  • backup_runs, app_settings — Google Drive backup history and operational settings such as the backup folder and retention (migration 0021). Only super admins can read them, and only the server’s backup job writes them, through the service role.
  • prompt_feed — a security_invoker view adding like/comment counts.
  • related_prompts(source_id, max_results) — a security invoker function that ranks prompts by shared tags, then title/note keywords (built-in full-text search), then likes and recency. It powers the related prompts on each prompt page.

All rules live in Postgres row-level security, so the browser-facing anon key is safe to ship:

  • A prompt is readable if it’s public, you own it, or you’re a member of its workspace.
  • Writes to a workspace prompt need the owner or editor role.
  • /api/ai-news reads no user data, so it skips the auth proxy and needs no key.
  • MCP calls go through security definer functions (mcp_search_prompts, mcp_get_prompt, mcp_create_prompt, mcp_list_tags) that resolve the token to a user first. No service-role key is needed anywhere.

GET /api/ai-news merges the latest AI posts into one list. It powers the in-app news feed and is public.

Source Where it comes from
OpenAI https://openai.com/news/rss.xml
Hugging Face https://huggingface.co/blog/feed.xml
Anthropic Hacker News stories linking to anthropic.com or claude.com (Anthropic publishes no RSS feed)
Groq Hacker News stories linking to groq.com (Groq publishes no RSS feed)
Community Hacker News stories with 20+ points mentioning LLM, OpenAI, Anthropic, Claude or Groq in the title

How it behaves:

  • Each upstream request and the merged result are cached for 15 minutes, so traffic to Prompy never reaches the sources directly.
  • A source that times out (8 s) or returns something that isn’t a feed is skipped and reported as ok: false. The other sources still load. The route returns 502 only when every source fails.
  • Items are de-duplicated by URL. When a vendor post is also on Hacker News, the vendor copy is kept.
  • Feed text is untrusted. Summaries are converted to plain text and cut to 280 characters, and only https: links and images are kept.
  • Admins can turn the whole feature off with the ai_news flag in /admin/flags. The route then returns 404.

Feeds rarely include images, and Hacker News has no summaries. So for the items on the page being served, and never the whole feed, Prompy reads each article’s own preview tags (og:image, og:description, twitter:*, meta description). The preview step:

  • caches each article’s preview for a day, so a page triggers at most limit fetches and repeat views trigger none
  • waits 3 seconds at most per page; slower previews finish in the background and show up on the next request
  • only fetches https: URLs on the default port whose hostname resolves to public addresses. IP literals, localhost, private, loopback and link-local ranges are refused, and each redirect is re-checked.
  • reads at most 256 KB of text/html per article, with a 4 s timeout, and keeps descriptions as plain text and images as https: URLs only
  • shows Hacker News points and comments in a separate meta field

Some sites, OpenAI among them, block automated requests. Their cards keep the feed’s own text and show the source instead of an image.

Param Default Allowed
page 1 1 or more
limit 10 1 to 30
source all all, anthropic, openai, groq, huggingface, community
{
"items": [
{
"id": "3f2a9c1b7e5d4a60",
"title": "Introducing …",
"summary": "Plain-text summary, at most 280 characters.",
"url": "https://openai.com/index/…",
"source": "openai",
"publishedAt": "2026-11-13T15:00:00.000Z",
"tags": ["product"],
"imageUrl": "https://…/cover.png",
"meta": "123 points · 45 comments on Hacker News"
}
],
"page": 1,
"limit": 10,
"total": 187,
"hasMore": true,
"sources": [{ "source": "openai", "ok": true, "count": 60 }]
}

imageUrl and meta are included only when there is one. Errors are { "error": "…" } with status 400 (bad parameters), 404 (feature off), 429 (more than 60 requests a minute from one IP) or 502 (every source down).

Admins switch features on and off at Admin → Flags (/admin/flags). A change applies on the next page load, with no deploy, and is recorded in Admin → Logs as feature_flag.updated.

The list of flags lives in code (web/lib/flags.ts), with a label, a description, a group and a default for each. The feature_flags table only stores the current value. If a flag has no row, or the table can’t be read, the default applies; switching such a flag in the admin page creates its row.

Flag Default What it controls
public_feed on The Explore feed (/explore) and its nav links. Off shows a paused notice.
courses_hub off The Learning Hub (/courses) and its nav link. Off returns 404.
ai_news on AI News & Blog (/blog), GET /api/ai-news and the landing page’s Latest in AI row.
app_bots on The AI chat section on the landing page.
pip_assistant on The Ask Pip chat. Off hides it and POST /api/chat returns 404.
ai_prompt_enhancer off Enhance prompt in the editor.
ai_auto_tags off Suggest tags in the editor.
image_search on Find an image in the editor (also needs an image provider key).
backups off The scheduled Google Drive backup. Back up now works either way.

To add a flag: add an entry to FLAGS in web/lib/flags.ts, seed its row in a new migration, and read it on the server with isEnabled("your_flag").