AI Radar — Your AI Intelligence Hub

An automated tracker for the global AI landscape: updates from 20 LLM labs, new models, coding agents, promos & free credits, AI startups, and new tools — in one fast, modern, mobile-friendly website.

Core principle: no fabricated data. Every item comes from a public source (official blogs/RSS, GitHub, Hugging Face, Hacker News, the OpenRouter API). When a source fails, the last valid data stays online and its health is shown transparently on /status.

Features

Page	What's inside
Today /	Command center. Overview: live ticker, AI Daily Briefing, Live Radar, cross-source Top stories, trending names/topics, "Reported, not yet official" safety net, bento (top story · buzz · new models · price moves). Workspace: sticky filters + lab list (left) · Global Feed (center, Cards/Compact) · sticky Social wire, Community pulse, Research radar, Deals (right)
AI Labs /labs	20 lab cards with logos, 30-day activity sparklines, "N new" badges since your last visit, coverage status
Lab detail /labs/[slug]	Full timeline (newest first): date/time, title, summary, category, practical benefit, source link, verification status, last-updated date; category filter, search, time range, pagination; per-source health
Discover /discover	Opportunity radar: X buzz (tips + posts linked by AI newsletters, read via X's official oEmbed), newsletters, new GitHub repos, Show HN, HF trending Spaces, Product Hunt (optional) — what it does, what's unique, benefit, pricing, why it's on the radar; separates brand new from older but trending; plus a Startup Radar of funding/launch headlines
Models /models	380+ models with live OpenRouter list prices, context length, free models, and a price-move tracker (every input/output price change, before → after)
Submit /submit	Send a tip (X post and/or product URL) → prefilled GitHub issue → published to Discover in minutes, bot replies on the issue
Deals /deals	Free tiers & official programs (re-verified every run with an evidence quote), free models live from OpenRouter (with end dates), lab promo announcements, community signals (flagged unverified)
Firehose /firehose	Every signal in one dense table — official, press, social (X & Threads), papers, community — with stream/lab/category/time filters, search, "strongest signal" sort, and CSV/JSON export
Digest /digest	The week in AI assembled automatically (top stories, models, pricing moves, tools & startups, research, social) — copy as Markdown
Saved /saved	Bookmarks (localStorage, no account); followed labs ("My Radar") are also stored locally
Status /status	Coverage per lab, health of every source, LLM status

Today also features a Live Radar (animated sweep; each blip is a real update from the last 72h, by lab) and a bento of the top story, X buzz, new models, and price moves.

Plus: Atom/JSON feeds (/feed.xml, /feed.json, /labs/<slug>/feed.xml), release-rhythm heatmap on /labs, follow labs (★), global search ⌘K / Ctrl+K / /, dark/light mode, read/unread & last-visited tracking (localStorage), SEO (metadata, sitemap, robots, OG image, JSON-LD).

Architecture

GitHub Actions (cron, every 3 hours)
  └─ npm run collect  ──►  data/*.json  ──► git commit ──► Vercel auto-deploy (static)
        │
        ├─ Labs: RSS/Atom · Hugging Face API · GitHub Releases (Atom) · new GitHub repos
        │     date normalization → cross-source dedupe → classification → ranking → summaries
        ├─ Discover: GitHub Search · Hacker News (Algolia) · Product Hunt* · watchlist
        ├─ Deals: official-page verification · OpenRouter models API · lab promo feeds · HN
        └─ Daily briefing (template; optional LLM*)                 * = optional via env

• Next.js 16 App Router (Cache Components), TypeScript, Tailwind CSS v4, shadcn/ui, Lucide.
• Every page is statically prerendered from data/*.json — fast and cheap (no database, no runtime API).
• Relative times ("3 hours ago"), unread state, and bookmarks are computed in the browser.

Quick start

npm install
npm run collect        # fetch real data (~1 min). Optional: GITHUB_TOKEN=... for higher rate limits
npm run dev            # http://localhost:3000

Script	Purpose
npm run collect	Full collector run (labs + discover + deals + briefing)
npm run collect -- --only=labs --lab=openai	One stage / one lab
npm run collect -- --backfill-days=90	Window for accepting items
npm test	Pipeline unit tests (parser, normalization, dedupe, classification)
npm run lint · npm run typecheck	Code quality
npm run build	Production build (also generates public/search-index.json)
npm run build:logos	Regenerate inline logos from public/logos/*.svg
npm run tip -- --x <post> --url <site>	Add a tip (X post / product) to src/config/tips.json
npm run summaries:pending / summaries:check	Editorial summaries queue / validator (daily Cursor run)

Environment variables

See .env.example.

Variable	Required	Notes
GITHUB_TOKEN	No (automatic in Actions)	Higher GitHub API limits for Discover
LLM_API_KEY	No	Enables richer summaries/analysis via a cheap LLM
LLM_BASE_URL	No	Any OpenAI-compatible endpoint (default OpenRouter)
LLM_MODEL	No	Default deepseek/deepseek-v4-flash
LLM_MAX_ITEMS	No	Max items summarized per run (default 40) — cost control
PRODUCT_HUNT_TOKEN	No	Enables Product Hunt in Discover
X_BEARER_TOKEN	No	Paid X API v2 recent search for launch posts (Discover works without it via tips + newsletters + oEmbed)
NEXT_PUBLIC_SITE_URL	No	Absolute URL for SEO (Vercel uses the production domain automatically)

API-frugal: the LLM only runs for new items (≤14 days) without an LLM summary, in batches of 8, and results are stored permanently in JSON — no item is ever summarized twice. Every number in LLM output must appear in the source text, otherwise the template is kept. Without an LLM, summaries use the source's own lead sentence plus deterministic templates.

Automation

• .github/workflows/collect.yml — cron 17 */3 * * * (every 3 hours), on new/edited tip issues (Discover only, so tips go live in minutes), plus a manual button (Actions → Collect AI updates → Run workflow). Runs tests, the collector, JSON validation, then commits data/ (with rebase/retry). Each commit triggers a Vercel deploy.
• .github/workflows/ci.yml — lint, typecheck, test, and build for PRs and code pushes.

Secrets/variables (Settings → Secrets and variables → Actions):

• Secrets: LLM_API_KEY, PRODUCT_HUNT_TOKEN, X_BEARER_TOKEN (all optional)
• Variables: LLM_BASE_URL, LLM_MODEL, LLM_MAX_ITEMS (optional)

Getting X buzz into AI Radar (no scraping)

X is never scraped. Posts reach AI Radar through legal paths only:

1. Tips — open /submit, paste the X post (and/or product URL) → a prefilled GitHub issue (label radar-tip). The issue triggers the collector, which reads the post via X's official oEmbed endpoint, resolves its links, loads the product page (name/description come from that page), and publishes it to Discover. The bot replies ✅ Added and closes the issue, or asks for info. Owner tips publish automatically; others need the approved label.
   • From Cursor/terminal: npm run tip -- --x https://x.com/user/status/123 --url https://product.com --note "why" (stored in src/config/tips.json).
2. Newsletters that curate AI Twitter (Ben's Bites) — their product links and linked X posts are followed automatically.
3. Optional X API — set X_BEARER_TOKEN for recent-search of launch posts.

Configuration: src/config/buzz.ts.

Daily editorial run (Cursor)

Once a day a scheduled Cursor agent acts as editor: it runs the collector, fixes broken sources, researches AI startups buzzing on X (added as "Agent pick" tips), and writes English summaries / "Why it matters" lines / the daily briefing into summaries/*.json. The site overlays those summaries at build time only if they validate (npm run summaries:check: known id, unchanged source, length limits, no URLs, every number present in the source). Copy-paste prompt: docs/CURSOR_DAILY_PROMPT.md.

Signals beyond official feeds (nothing slips through)

Stream	Source	How it is labeled
Official	Lab blogs/RSS, GitHub Releases, Hugging Face	verified when the link is on the lab's own domain/org
Press	Google News per lab (20 queries), grouped by story	never verified; multi-outlet reports with no matching official post surface as "Reported, not yet official"
Papers	Hugging Face Daily Papers, attributed to a lab by organization	—
Community	Hacker News (AI stories ≥ 60 points)	—
Social	X and Threads — official access only (below)	posts from a lab's own accounts are marked "official"

Cross-source stories (data/stories.json) group items about the same model/product (e.g. "Claude Haiku 5.5" across Anthropic, AWS, and media) by primary entity — never chained through comparison headlines. Price moves are debounced (a new price must hold for 3 consecutive snapshots) and router aliases are ignored.

X and Threads

X and Threads are never scraped. Posts enter through:

1. oEmbed (X's official embed endpoint) for posts linked from curated newsletters (Ben's Bites) and from tips — including posts by the labs' own accounts (LAB_SOCIAL in src/config/signals.ts).
2. Tips: /submit (X or Threads URL → GitHub issue) or npm run tip -- --x <url> --by cursor. Threads posts are stored link-only (Threads oEmbed needs an app token).
3. X API v2 (X_BEARER_TOKEN, paid X developer access): recent posts from every official lab handle, with metrics.
4. Threads API (THREADS_ACCESS_TOKEN): keyword search + profile posts for configured handles.

Without the two tokens, X/Threads coverage is what curators and tips surface — not every post. Add the tokens (GitHub → Settings → Secrets → Actions) to ingest official accounts automatically; /status shows which channels are active.

Source configuration

All sources live in one place and are easy to change:

• src/config/labs.ts — 20 labs + per-lab sources (RSS, Hugging Face, GitHub Releases, new GitHub repos), include/exclude filters, official domains for verification.
• src/config/discover.ts — GitHub queries, HN thresholds, repo watchlist.
• src/config/deals.ts — free-tier/credit programs + required evidence text (mustContain).

Full guide (Indonesian): docs/SOURCES.md. Daily maintenance with Cursor (Indonesian): docs/CURSOR_GUIDE.md.

Current coverage

Coverage	Labs
Official feed (official RSS/blog + GitHub/HF)	OpenAI, Google DeepMind, Microsoft, Amazon, NVIDIA, IBM
Via mirror (no RSS → community mirror of official pages + official GitHub/HF)	Anthropic, xAI, Meta, Mistral, Cohere
Partial (no machine-readable blog → official Hugging Face + GitHub)	DeepSeek, Qwen, Moonshot Kimi, Z.ai GLM, MiniMax, Baidu, Tencent, ByteDance, AI21

X/Twitter is never scraped. Labs with nothing new in 90 days show their latest real releases (up to 1 year) so all 20 cards always have content.

Project structure

src/app/            pages (Today, labs, discover, deals, saved, status) + sitemap/robots/OG
src/components/     UI (update cards, timeline, lab grid, ⌘K search, filters, theme)
src/config/         source registry: labs.ts, discover.ts, deals.ts
src/lib/            data types, server data loader, formatting, localStorage state
scripts/collect.ts  collector entrypoint
scripts/collector/  http, feed parser, sources, merge/dedupe, classify, summarize, llm, discover, deals, briefing
data/               collector output (committed): updates/<lab>.json, discover.json, deals.json, briefings.json, status.json
tests/              pipeline unit tests (vitest)

Deploying to Vercel

1. Import this repo in Vercel (framework: Next.js, no extra config).
2. Optionally set NEXT_PUBLIC_SITE_URL for a custom domain.
3. Every data commit from GitHub Actions triggers a redeploy automatically.

License & attribution

Lab logos from @lobehub/icons (MIT) — trademarks belong to their respective owners. Community RSS mirrors: Olshansk/rss-feeds.