InstaPilot AI

An AI worker that runs your Instagram and YouTube Shorts channels while you sleep.

It picks the topic, writes the post, designs the slide cards, renders a real vertical video with an AI voiceover, publishes it on your schedule, cross-posts it as a Reel, and answers the comments and DMs it gets back. Any niche, your accounts, your API keys.

▶ Try the live demo — the real dashboard running a fictional brand in your browser. Nothing is posted and no AI is called.

_{Not a mockup. These are frames from a real Short — topic, script, cards, voiceover, video, title, caption and tags all generated and published unattended.}

---

What it actually does

Every day, for each brand you configure, unattended:

1. Writes an on-topic post with AI — rotating through your topics, and generating fresh ones in the same style when they run out. It de-duplicates at the theme level, not just the wording, because platforms suppress repetitive uploads.
2. Designs branded image cards — posts, multi-slide carousels, Stories — rendered server-side with Satori and Sharp.
3. Renders a vertical Short from those cards with ffmpeg: a ~2s curiosity hook cover (also uploaded as the custom thumbnail), large-text content slides, and a subscribe outro — with mood-matched Creative Commons music, an optional AI voiceover synced card-by-card, and optional word-by-word burned captions.
4. Publishes to YouTube and Instagram on a per-weekday schedule you set, then cross-posts the Short to Instagram as a Reel on its own deferred timing.
5. Replies to Instagram comments and DMs (including voice notes — Whisper in, TTS out) and to YouTube comments, in whatever language the person used.
6. Reports — daily health email, morning digest, live SSE alerts, real analytics pulled from the Graph and Data APIs.

See it running

The pipeline runs a live channel end to end — topic, script, cards, video, title, caption, tags and upload:

• The Cardio Doc AI — 105 Shorts written, rendered and published on schedule.
• Example Short — "How a 6-Second Breath Can Drop Your Heart Rate", 33s, hook cover → four content slides → subscribe outro (the frames above).
• The Instagram side — the same renderer's output cross-posted as a Reel, carrying the AI-written rich caption: hook, six numbered stat-backed points, then the dual-account follow CTA the caption builder assembles.

To be straight about it: this is proof the machine runs, not proof it will make you famous. The channel is small. What 105 unattended uploads demonstrate is that the scheduler, the renderer, the caption chain and the upload path hold up day after day without a babysitter.

Why it might be interesting

• It renders the video itself. No Canva, no template service, no headless browser. Cards are laid out with Satori, rasterised with Sharp, stitched to a 720×1280 H.264 MP4 with ffmpeg, then mixed with music and narration. See lib/videoGenerator.ts.
• Per-task AI fallback chains. Content, reply, and vision each get their own ordered provider · model chain across Groq, Cerebras and Gemini, tried top-to-bottom. One provider's outage or quota wall falls through instead of failing the day.
• Genuinely white-label. App name, niche, persona, handles, content-type labels and prompts are all configured in Settings — nothing niche-specific is compiled in. There is an AI Setup wizard that interviews you and writes the whole config.
• Multi-brand. Each brand is a paired IG account + YouTube channel with its own skin, schedule, topics and prompts. The engine runs the full pipeline independently for every active brand.
• Bring your own keys. Nothing is bundled. Groq, Gemini, Meta, Cloudinary, Resend, and optionally YouTube and Jamendo — all yours.

Quickstart

git clone https://github.com/ys941/instapilot-ai && cd instapilot-ai
npm install
cp .env.example .env                # fill in your keys
docker compose up -d postgres       # or point DATABASE_URL at your own database
npm run db:generate && npm run db:push
npm run dev                         # http://localhost:3000

You do not have to invent a login key. On the first run, .env is created for you from .env.example with a working DATABASE_URL, a random SESSION_SECRET, and a random APP_ACCESS_KEY that is printed in the terminal — change it in .env whenever you like. The AI features stay switched off until you add your own API keys.

Use .env, not .env.local. The Prisma CLI reads .env only — it never reads .env.local, which Next.js prefers at runtime. Keeping your config in a single .env avoids db:push writing the schema to one database while the running app reads another. If you do keep a .env.local for overrides, DATABASE_URL must be identical in both. npm run dev and npm run db:push both run npm run check-env first, which explains anything still missing instead of failing later with an opaque Prisma or login error.

Log in with your APP_ACCESS_KEY, then open Settings → AI Setup and describe your brand in plain English — it generates the skin, content types, topics, schedule and persona for you. Or configure each tab by hand.

Prefer Docker:

cp .env.example .env && docker compose up -d

On Windows, start-all.bat brings up Postgres, Redis, Prisma and the dev server in one step.

Screenshots

_{The dashboard, on a fresh install before any brand is connected.}

Overview — health, today's queue, live activity	AI Generator — pick platform, type and tone
Scheduler — per-weekday plan, per-platform	Analytics — live IG + YouTube stats

More — brand skin, AI config, auto-post, digest

How it works

A single in-process loop (lib/catchup.ts) drives everything — started by instrumentation.ts on boot, re-fired every 5 minutes, plus a daily timer and Instagram webhooks. Each cycle it lists every active brand and runs the full pipeline for each one independently. Work is idempotent and self-healing: a crashed render or a provider outage is retried on the next pass rather than silently dropping the day.

Stack: Next.js 16 · TypeScript · PostgreSQL + Prisma · Tailwind · Satori + Sharp · ffmpeg-static · Groq / Cerebras / Gemini · Facebook Graph API · YouTube Data API v3 · Cloudinary · Railway or Docker.

Documentation

Features	Every capability, in detail
Configuration	All settings and environment variables
Architecture	Engine, routing, project structure
Deployment	Railway, Docker, webhook setup
Security	Auth model, HMAC, SSRF guard
Contributing	Five-minute setup and open issues

Status and caveats

This is a working system that publishes to real accounts, not a demo. Some things you should know before you run it:

• You are responsible for what it publishes. Meta and YouTube both have automation policies; read them. Set the schedule conservatively.
• The Shorts pipeline is CPU-heavy. Voiceover and burned captions are opt-in because they add a TTS call, a Whisper call and a re-encode per Short.
• Vision captions, hashtags and YouTube tags currently resolve the primary brand's vision chain regardless of which brand is publishing.
• There is no hosted demo, because a demo would have to publish to a real Instagram account. The live channel linked above is the demo.

Licence

Standard MIT — use it, fork it, re-skin it, run client channels on it, keep the money. Alongside it the project asks that credit to the original author stays visible: the dashboard footer links to @ys941 (the app name above it stays fully white-label), and the app checks at startup that ATTRIBUTION_ACK="https://github.com/ys941" is set in your environment. Nothing is transmitted — the value is compared locally. See COPYRIGHT.md and lib/attribution.ts.

All third-party API usage remains subject to those providers' terms.

Contributing

Contributions are genuinely welcome, including your first one. CONTRIBUTING.md has a five-minute local setup, a map of the codebase, and a list of ideas looking for an owner.

Star the repo

If InstaPilot is useful to you, star it. There is no marketing behind this project — a star is genuinely how the next person ends up finding it.