Cut through the noise. Surface only what's worth acting on.
Signal is a self-hosted, personalized news filtering app. It scrapes RSS feeds, Reddit, and Hacker News daily, then runs every story through Claude Haiku against your scoring profile — so your feed surfaces only the high-signal content that matters to you.
Live demo: signal-nu-nine.vercel.app
- Scrapes 20+ RSS feeds, Reddit subreddits, and Hacker News daily (free, no API keys needed for scraping)
- Filters every story through Claude Haiku with your personal scoring profile — not a generic relevance algorithm
- Surfaces a ranked feed by category: Opportunities, Ideas, and Intel
- Notifies you via Web Push when a high-score story drops
- Costs ~$0.50–$3/month to run (you bring your own Anthropic key; Haiku only in the filter pipeline)
| Landing | Feed |
|---|---|
 |
 |
[Daily Cron] → /api/scrape → RSS + Reddit + HN + Nitter
↓
raw_stories (shared pool, Supabase)
↓
[User: Run Pipeline] → decrypt BYOK key → Claude Haiku (batch, 24 stories/call)
↓
scored_stories (per user_id)
↓
/feed → ranked by score, filtered by category
↓
Web Push → high-score alerts
User setup flow: Sign up → verify email → add Anthropic key in Settings → complete 10-question onboarding (Claude Sonnet synthesizes a scoring profile) → /feed
| Layer | Technology |
|---|---|
| Frontend / Backend | Next.js 16 (App Router) |
| Database + Auth | Supabase (Postgres + RLS) |
| AI Filtering | Claude Haiku (claude-haiku-4-5) — batch scoring only |
| Profile Synthesis | Claude Sonnet — one-shot at onboarding |
| Styling | Tailwind CSS |
| Hosting | Vercel (Hobby tier) |
| Push Notifications | Web Push (VAPID) |
- Supabase account (free tier works)
- Vercel account (Hobby tier works)
- Anthropic API key (each user brings their own — BYOK)
git clone https://github.com/KBKote/signal.git
cd signal
npm install- Create a new Supabase project
- Run the full schema in the SQL editor:
# Copy the contents of supabase/schema.sql and run in Supabase SQL editor- Apply migrations in order (Supabase SQL editor or
supabase db push):
supabase/migrations/20260411120000_scrape_user_throttle.sql
supabase/migrations/20260411130000_filter_user_throttle.sql
supabase/migrations/20260411140000_api_scored_stories_page.sql
supabase/migrations/20260411150000_prune_signal_story_tables.sql
supabase/migrations/20260412100000_scoring_markdown.sql
supabase/migrations/20260412200000_atomic_rate_limit.sql
supabase/migrations/20260412210000_auth_rate_limit.sql
supabase/migrations/20260422110000_fix_ivfflat_probes.sql
- In Supabase Auth settings:
- Enable Email provider with email + password
- Set Site URL to your production URL
- Add
/auth/callbackto Redirect URLs
cp .env.local.example .env.localFill in .env.local:
# Required
NEXT_PUBLIC_SUPABASE_URL=https://your-project.supabase.co
NEXT_PUBLIC_SUPABASE_ANON_KEY=eyJ...
SUPABASE_SERVICE_ROLE_KEY=eyJ...
# AES-256-GCM key for encrypting user Anthropic keys at rest
SECRETS_ENCRYPTION_KEY= # openssl rand -base64 32
# Sonnet model for one-shot profile synthesis at onboarding
ANTHROPIC_PROFILE_MODEL=claude-sonnet-4-6
# Web Push (generate with: npx web-push generate-vapid-keys)
NEXT_PUBLIC_VAPID_PUBLIC_KEY=...
VAPID_PRIVATE_KEY=...
VAPID_SUBJECT=mailto:you@example.com
# Production: protects the /api/scrape cron endpoint
CRON_SECRET=
NEXT_PUBLIC_SITE_URL=https://your-app.vercel.app
# Optional: in-app bug reports (Nodemailer SMTP — e.g. iCloud smtp.mail.me.com:587 + app-specific password)
# BUG_REPORT_TO_EMAIL=
# BUG_REPORT_FROM_EMAIL=
# SMTP_HOST=
# SMTP_PORT=587
# SMTP_USER=
# SMTP_PASS=npm run dev # http://localhost:3000
npm run tunnel # Cloudflare quick tunnel (for OAuth callbacks)vercel deployAdd all environment variables in Vercel's project settings. The daily scrape cron (vercel.json) runs automatically at midnight UTC on Vercel Hobby.
Signal is designed to be cheap to run. Each user brings their own Anthropic key (BYOK), so your operator costs are near zero.
| Item | Cost |
|---|---|
| Vercel Hobby | Free |
| Supabase Free tier | Free |
| RSS / Reddit / HN scraping | Free |
| Claude Haiku (per user, per daily run) | ~$0.01–0.05/run |
| Claude Sonnet (one-shot onboarding only) | ~$0.01 one-time |
| Typical monthly per user | < $1.50 |
Haiku costs are kept low by: keyword pre-filtering before any Claude call, batching 24 stories per API call, and capping why to 12 words and summary to 20 words in the response.
Edit lib/scrape-sources.ts to add or remove:
- RSS feeds — any feed URL
- Subreddits — just the subreddit name
- Nitter accounts — Twitter/X accounts via free Nitter RSS
The scoring profile is generated at onboarding from your answers to 10 questions. Claude Sonnet synthesizes a scoring_markdown document that Haiku uses on every filter run. You can view and edit your profile in /settings.
signal/
├── app/ # Next.js App Router
│ ├── page.tsx # Landing page
│ ├── feed/ # Main feed (auth-gated)
│ ├── settings/ # BYOK key + scoring profile editor
│ ├── onboarding/ # 10-question profile setup
│ └── api/
│ ├── scrape/ # Trigger scraper (cron + manual)
│ ├── filter/ # Run Claude Haiku pipeline (user BYOK)
│ ├── stories/ # Fetch scored feed with pagination
│ └── onboarding/ # Synthesize scoring profile (Sonnet)
├── lib/
│ ├── scraper/ # RSS, Reddit, HN, Nitter collectors
│ ├── filter.ts # Claude Haiku batch scoring pipeline
│ ├── auth/ # Gate logic (email → BYOK → profile → feed)
│ ├── crypto-user-secrets.ts # AES-256-GCM encrypt/decrypt for BYOK
│ ├── pipeline-preferences.ts # Scope settings (precise/balanced/expansive)
│ └── scrape-sources.ts # All RSS feeds, subreddits, Nitter accounts
├── components/ # React UI components
├── supabase/
│ ├── schema.sql # Full DB schema + RLS policies
│ └── migrations/ # Incremental SQL migrations
├── proxy.ts # Auth routing (replaces Next.js middleware)
├── .env.local.example # Environment variable template
└── vercel.json # Cron schedule (daily scrape)
- BYOK isolation: Each user's Anthropic key is encrypted at rest with AES-256-GCM using a server-side
SECRETS_ENCRYPTION_KEY. The key is decrypted only at request time, never logged, and never used as a fallback for other users. - Row-Level Security: All Supabase tables use RLS policies. Users can only read/write their own rows.
- Gate chain: Every protected API route enforces session → email verified → BYOK key → scoring profile in order.
- Rate limiting: DB-backed atomic rate limits on scrape (per user) and filter (90s cooldown per user).
- Vercel Hobby cron: One cron per day max — scrape runs once at midnight UTC. If you need more frequent scraping, move
/api/scrapeto a separate worker (Railway, Fly.io, or Cloudflare Worker). - Pool depletion: After multiple filter runs in a day, the unscored candidate pool empties. Run the scraper first to replenish.
- Nitter reliability: Public Nitter instances are unreliable.
breakdown.nitter === 0in scrape logs usually means the instance was down, not a code bug. - Memory on Vercel:
next buildmay be killed in low-memory environments. The app is known to work fine on standard Vercel deployments.
MIT — see LICENSE.
Pull requests are welcome. Before submitting:
npm run lint # must pass
npm run build # must passFor significant changes, open an issue first to discuss the approach. Keep the filter pipeline on Haiku only — Sonnet/Opus would break the cost model for self-hosters.