MyNews — Vazghen's news feed

A multi-topic news aggregator that doubles as an API your AI agents can read. One small Express server pulls RSS for 16 topics, serves a designed browser feed, and exposes the same data as a REST + MCP API — including full readable article text, not just headlines. No API keys required.

Live: https://mynews-nine.vercel.app · Repo: https://github.com/iamvazghen/mynews

/                 landing page
/dashboard.html   the live feed — filter, sort by newest/category, timeframe, search
/api/*            the agent API (REST), also exposed over MCP via mcp.js

Why I built it (portfolio note)

A focused full-stack project that solves a real problem I have: I read across ~16 unrelated topics, and I want my own agents to read them too. It demonstrates:

• Full-stack JS — Express API, RSS aggregation, in-memory caching, HTML→text article extraction.
• Design — a hand-built Cobalt design system (OKLCH tokens, Space Grotesk / Inter / JetBrains Mono), responsive, reduced-motion-safe; no UI framework.
• Agent integration — the same API is consumed by a browser and by LLM agents over MCP, so a scheduled agent can fetch headlines and then read full articles to summarize.
• Shipping — deployed to Vercel (serverless) from GitHub with CI on push, and to a private VPS as a systemd service.

It's deliberately small: ~1 backend file, 1 feed-config file, a static frontend, and an MCP shim.

Run it locally

npm install
npm start          # → http://localhost:3000
npm test           # unit checks for reading-time + filter/sort

The feed is live immediately — every card shows when it dropped and how long it takes to read; sort by newest or category, pick a timeframe, search.

The live feed — what's wired and how to change it

Everything goes through `feeds.js` — the only file you edit. Each category is { slug, label, feeds: [rss urls] }. A dead/blocked URL is silently skipped, so the page never breaks; swap it when convenient.

• Add a source: drop another RSS URL into a category's feeds array.
• Add a category: append a new { slug, label, feeds } object. Frontend, API and chips pick it up automatically.
• Find a feed: most sites expose /feed/, /rss, or /feed.xml. Topics without first-party RSS use Hacker News query feeds (hnrss.org/newest?q=...).

Optional API keys (none required)

RSS covers all 16 topics today, so nothing is needed to run or deploy. Add a key only if you want more than a source's feed exposes — each is a small change:

Store keys as env vars (add dotenv) — never hardcode them.

The API

CORS is open, so any agent or page can call it.

GET /api/categories
  → [{ slug, label }, ...]

GET /api/news?category=ai,crypto&since=24&sort=newest&q=bitcoin&limit=50
  category  comma-separated slugs (omit = all)
  since     only items from the past N hours (0 = all time)
  sort      newest | category
  q         keyword filter on title + snippet
  limit     max items (default 200, cap 500)
  → { count, totalReadingMinutes, items: [{ title, link, source,
       category, publishedAt, readingMinutes, snippet, image }] }

GET /api/article?url=<article-url>
  → { url, title, author, published, source, wordCount, readingMinutes, content }
    full readable body text, boilerplate stripped — what an agent reads to summarize a story

curl "https://mynews-nine.vercel.app/api/news?category=agents,ai&since=24&limit=10"
curl "https://mynews-nine.vercel.app/api/article?url=https://www.theverge.com/some-story"

Connect your AI agent — step by step

The API is public and keyless, so wiring any agent takes three steps.

1. Point at the API. Base URL https://mynews-nine.vercel.app/api.
2. Get headlines. GET /api/news?category=ai,crypto&since=24 → titles, sources, timestamps, read-time.
3. Read the full story. GET /api/article?url=<link from step 2> → clean body text to summarize.

That's enough for any framework (LangChain, a custom tool, a cron + curl). For MCP-native agents, use the bundled server instead of raw HTTP:

npm i @modelcontextprotocol/sdk zod
NEWS_API=https://mynews-nine.vercel.app node mcp.js   # speaks MCP over stdio

It exposes three tools: get_news, get_article, list_categories.

OpenClaw (the simplest path)

Add one block to ~/.openclaw/openclaw.json under mcp.servers, then restart the gateway — every agent gains the three tools:

"mynews": {
  "command": "/usr/bin/node",
  "args": ["/home/openclaw/newsfeed/mcp.js"],
  "env": { "NEWS_API": "https://mynews-nine.vercel.app" }
}

A morning-news cron then calls get_news for the day's headlines and get_article for the full text of each story it wants to summarize — no scraping glue.

Deployments

• Vercel (public, online): Express runs as a serverless function (api/index.js + vercel.json); public/ is the static site. Auto-deploys on push to main.
• VPS (private): a systemd --user unit (newsfeed.service, Restart=always) on port 3000. Manage with systemctl --user restart|status newsfeed and journalctl --user -u newsfeed -f.

Files

• server.js — API + cache + reading-time + article extraction
• feeds.js — the 16 categories → RSS URLs (edit this)
• public/ — Cobalt-themed landing + feed (index.html, dashboard.html, app.js, styles.css, landing.css, tokens.css)
• api/index.js + vercel.json — Vercel serverless wiring
• mcp.js — MCP server (get_news / get_article / list_categories)
• test.js — unit checks