title	Newsletter Digest API
emoji	🚀
colorFrom	blue
colorTo	indigo
sdk	docker
pinned	false

Newsletter Digest

Pull recent newsletters from your Gmail, AI-filter them, then generate three text outputs:

1. Digest — themed summary of every newsletter in the window
2. Story — narrative version of the digest
3. LinkedIn post — short post derived from the story

Single-user: no auth, just don't share the URL. Email + IMAP password + lookback hours + prompts persist in your browser. Past runs persist in Supabase, scoped by the email you entered.

Stack

• Backend — FastAPI on Render (backend/)
• Frontend — Next.js on Vercel (frontend/)
• DB — Supabase Postgres (one runs table)
• LLM — OpenRouter, model meta-llama/llama-3.3-70b-instruct:free

0. Prereqs

• A Gmail account with an App Password (16 chars). Generate at https://myaccount.google.com/apppasswords. You need 2-Step Verification on first.
• Accounts on: Supabase, Render, Vercel, OpenRouter.

1. Supabase

1. Create a project. Note the Project URL and the service_role key (Settings → API). The service key is secret — never put it in the frontend.
2. Open the SQL Editor and run the contents of supabase_schema.sql.

2. Backend on Render

This repo includes render.yaml — Render will pick it up automatically.

1. Push the repo to GitHub.
2. Render → New → Blueprint, point at the repo. It will detect render.yaml and create the newsletter-digest-api web service.
3. Set these Environment variables on the service:
   • OPENROUTER_API_KEY — your OpenRouter key
   • SUPABASE_URL — https://xxxx.supabase.co
   • SUPABASE_SERVICE_KEY — the service_role key from step 1
   • ALLOWED_ORIGINS — your Vercel URL (e.g. https://newsletter-digest.vercel.app). Use * only while testing.
4. Deploy. After it's up, hit https://<your-service>.onrender.com/ — should return {"ok": true, ...}.

Render free tier sleeps after ~15 min idle. The first request after a sleep takes ~30s to wake. Subsequent runs are fast.

3. Frontend on Vercel

1. Vercel → Import Project → pick the repo.
2. Root Directory: frontend
3. Environment variable:
   • NEXT_PUBLIC_API_BASE = https://<your-service>.onrender.com
4. Deploy.

After Vercel gives you a URL, go back to Render and set ALLOWED_ORIGINS to that URL, then redeploy the backend so CORS picks it up.

4. Use it

Open the Vercel URL.

1. Enter your Gmail address and 16-char app password.
2. Click Verify connection — confirms IMAP login works.
3. Set hours-back (default 24). Optionally edit prompts under "Prompts (advanced)".
4. Click Run. Spinner runs while the pipeline:
   • fetches mails from the last N hours across all folders
   • asks the LLM (one batched call) which are newsletters
   • aggregates the kept bodies
   • generates digest → story → LinkedIn post
5. Results show in tabs. Past runs appear at the bottom and persist in Supabase.

Daily auto-posting

After the Make.com LinkedIn webhook is working manually:

1. In Settings, set the lookback window, then enable Auto-post daily.
2. Pick the daily post time and timezone.
3. Click Save Settings.
4. Configure the Render cron job from render.yaml with:
   • AUTOMATION_TICK_URL = https://<your-service>.onrender.com/automation/tick
   • CRON_SECRET = the same value as the backend service, if you set one

The cron job pings the backend every 5 minutes. The tick endpoint checks all enabled users, runs only users whose local scheduled time has passed, uses each user's saved lookback window, and records last_run_at so each user runs at most once per local day.

Optional: set CRON_SECRET on the backend and send it as the X-Cron-Secret header from your scheduler. If CRON_SECRET is not set, the tick endpoint is open but still only runs saved schedules that are due.

Local development

# Backend
cd backend
python -m venv .venv && source .venv/bin/activate   # Windows: .venv\Scripts\activate
pip install -r requirements.txt
cp .env.example .env   # fill in keys
# load env then run:
uvicorn main:app --reload --port 8000

# Frontend (separate terminal)
cd frontend
npm install
echo "NEXT_PUBLIC_API_BASE=http://localhost:8000" > .env.local
npm run dev

Notes

• The OpenRouter free model has rate limits. Each run makes 4 LLM calls (filter, digest, story, LinkedIn). The backend retries on 429/5xx with exponential backoff.
• All errors surface verbatim to the UI (red banner). No silent fallbacks.
• Mail bodies > 8000 chars are truncated per-mail; the aggregate is capped at 80k chars to fit context.
• "Newsletter" detection is purely the filter prompt — edit it in the UI to taste.