A self-hosted, zero-cost trend blog. A scheduled job runs every hour, picks the highest-signal story from seven sources, researches it, writes a structured MDX post, and commits it to GitHub. The Next.js site can redeploy from those commits using the host configured for the repository.
Stack: Next.js 15 · TinaCMS · Groq (free tier) · Brave Search · Pexels · GitHub Contents API · Cloudflare/self-hosted deployment.
Monthly cost at steady state: $0 when the selected services remain within their free tiers.
┌─ Reddit ─┐
│ HN │
│ DEV.to │──▶ score ──▶ dedup ──▶ winner ──▶ research ──▶ Groq ──▶ MDX ──▶ git commit ──▶ deploy
│ RSS │ (pop + engagement + recency) (Brave + scrape (strict JSON
│ YouTube │ + YT transcripts) contract)
└─ Brave ─┘
Each stage is its own module in src/lib/orchestrator/ and can be tested independently. The pipeline.ts runner wires them together with per-stage timings and graceful fallbacks. A flaky source does not kill the run.
npm (the repo ships package-lock.json; CI uses npm ci)npm install
cp .env.example .env.local
| Key | Where | Free tier |
|---|---|---|
GROQ_API_KEY |
https://console.groq.com/keys | Generous rate limits, ~30 RPM on llama-3.3-70b |
BRAVE_API_KEY |
https://api.search.brave.com/app/keys | 2,000 queries/month on the free plan |
PEXELS_API_KEY |
https://www.pexels.com/api/new/ | Unlimited for dev use |
GITHUB_TOKEN |
github.com → Settings → Developer settings → Fine-grained PAT | Scope: Contents: Read/Write on the blog repo only |
CRON_SECRET |
openssl rand -hex 32 |
n/a |
Fill them into .env.local along with GITHUB_OWNER / GITHUB_REPO / GITHUB_BRANCH.
Security note: Never commit
.env.localor any file containing real API keys to version control. The.env.localfile is already in.gitignoreto prevent accidental commits. Always use.env.exampleas a template with placeholder values only.
# Dry run: prints the generated post, does not write anything
npm run generate -- --dry
# Real run: writes MDX to content/posts/ and updates content/.topic-log.json
npm run generate
# Start the dev server
npm run dev
Open http://localhost:3000. The seed post is visible out of the box; new posts show up as soon as npm run generate writes them.
The hourly schedule lives in .github/workflows/generate.yml, which runs at the top of every hour (cron: '0 * * * *'), executes the pipeline with npx tsx scripts/run-local.ts, and commits any new post straight to the repo. The push can then trigger whichever hosting path is configured for the project.
Add the pipeline secrets (GROQ_API_KEY, BRAVE_API_KEY, PEXELS_API_KEY, REDDIT_CLIENT_ID, REDDIT_CLIENT_SECRET) under Settings → Secrets and variables → Actions. The workflow has contents: write and a concurrency group so a slow run never overlaps the next tick. Use the Run workflow button (workflow_dispatch) to trigger a one-off run.
Keep scheduling in one place. If you later attach another scheduler to /api/cron/generate, disable the GitHub Actions tick first so the pipeline cannot generate duplicate posts.
If the repository’s current Next.js output is compatible with the selected Pages setup, use Cloudflare for the site while leaving content generation in GitHub Actions. Do not move the 30–90 second generation pipeline into an edge function with a shorter CPU budget.
npm run build && npm start and point a reverse proxy at port 3000. The GitHub Action still drives generation; to trigger the route manually, use your configured public origin:
curl -H "Authorization: Bearer $CRON_SECRET" https://your-domain/api/cron/generate
No production hostname is hard-coded by this repository.
The schema in tina/config.ts matches the frontmatter the pipeline emits. Start the editor with:
npm run dev # Tina runs alongside Next via the `tinacms dev` wrapper
Then visit http://localhost:3000/admin/index.html. You can fix typos, tweak tags, or hand-write posts that follow the same structure.
Self-hosted mode (default): TinaCMS works in local filesystem mode without any cloud credentials. The build script (scripts/build.sh) automatically handles this by setting placeholder values during build if credentials are not provided.
Hosted editing: For non-local contributors, sign up at tina.io for the free tier and fill in NEXT_PUBLIC_TINA_CLIENT_ID + TINA_TOKEN in your deployment environment variables. These are optional for local development.
Every generated post follows this exact shape. The system prompt in src/lib/orchestrator/generate.ts enforces it, and the zod schema validates the JSON before writing:
<Callout type="takeaway"> with a one-sentence synthesis## What happened## Why it matters<ProsCons> block with 3+ items per side## How to think about it<Callout type="warning"> when warranted## FAQ with exactly 3 <Question> entriesAll components are implemented in src/components/mdx/index.tsx and styled via globals.css’s .prose-editorial rules.
From src/lib/orchestrator/score.ts:
score = 0.5·popularity + 0.2·engagement + 0.3·recency
Dedup uses a sorted-token fingerprint of the title. The topic log (content/.topic-log.json) is checked on every run and capped at 500 entries.
“no items from any source”: all sources failed. Check logs and retry after the upstream services recover.
“all top candidates already covered”: the scorer found winners, but their signatures already exist in the topic log.
“no research content scrapable”: the winner’s URL and the research fallbacks could not be scraped. The pipeline skips gracefully.
Groq rate limit: the free tier resets on its own cadence. One post/hour is intentionally conservative.
Hosting timeout: keep the long-running generation pipeline in GitHub Actions rather than moving it into a short-lived edge request.
src/lib/sources/, export a function returning RawItem[], and add it to the Promise.all in pipeline.ts.SYSTEM_PROMPT in generate.ts. The zod schema catches structurally invalid output.SUBREDDITS in reddit.ts, BRAVE_QUERIES in bravenews.ts, and DEFAULT_FEEDS in rss.ts.cron in .github/workflows/generate.yml.MIT.