ShortSmash

ShortSmash ⚒️

Prompt in. Finished short out.

ShortSmash is an AI video studio that takes one sentence — or today's trends — and ships a finished, captioned 9:16 short, then publishes it to YouTube, TikTok, Instagram, X, Threads, Bluesky, LinkedIn, Pinterest, and Facebook (profiles + Pages) through their official APIs. Script, voiceover, *verified* stock footage, karaoke captions, music ducking, trend harvesting, batch production, scheduling and comment-reply drafts, all in one CLI. And an MCP server, so AI agents can drive the whole factory.

shortsmash trends food                    # what's hot right now (free, no keys)

shortsmash batch food --count 5 # turn the top 5 trends into 5 videos shortsmash serve # 🆕 web studio: http://localhost:8000 shortsmash agent "make a 30s reel about coffee and post it Friday at 6pm" shortsmash post myvideo.mp4 --caption # post YOUR video: AI title + hashtags shortsmash post myvideo.mp4 --at "Friday 6pm" --platforms youtube,tiktok

git clone https://github.com/hydope/shortsmash && cd shortsmash

./install.sh # Windows: .\install.ps1 shortsmash "why you can't tickle yourself" # prompt in, finished short out

🏆 Featured in

ShortSmash is listed in curated awesome lists (merged PRs):

• ToolkitlyAI/awesome-ai-tools — AI Video
• TensorBlock/awesome-mcp-servers — MCP registry profile live on tensorblock.co
• sajadh76/awesome-AI-tools — Video Generation
• mcpHQ/awesome-mcp-servers — MCP Servers
• Jenqyang/Awesome-AI-Agents — AI Agents
• Official MCP Registry — io.github.hydope/shortsmash, installable via pip install shortsmash

🏗️ Architecture

flowchart LR

P["Prompt / trend"] --> LLM["LLM: script + shot plan (NVIDIA · Gemini · Grok)"] LLM --> TTS["Neural TTS (50 languages)"] LLM --> VIS["Visuals (Pixabay · Agnes · Veo 3.1 · Kling)"] VIS --> GATE["Vision gate (CLIP + Gemini)"] GATE -->|reject| VIS TTS --> REN["FFmpeg render 1080×1920 · karaoke captions cinematic grade · music"] GATE --> REN REN --> DUB["Dub engine (50 languages)"] REN --> PUB["10 publishers (official APIs)"] DUB --> PUB PUB --> SCHED["Scheduler (SQLite · 'Friday 6pm')"]

🖥️ ShortSmash Studio (web UI)

shortsmash serve launches a dark, fast studio in your browser — no build step:

• 🎬 Studio tab — type a topic, hit Smash, watch it build; video previews inline
• 📈 Trends tab — browse what's hot per niche, click any trend to use it as your topic
• 📚 Library tab — every project, video preview, one-click publish
• 📤 Manual posts — upload your own video, get a pre-flight check (resolution, duration, per-platform limits), AI-written title + hashtags, then post now or schedule it ("Friday 6pm") to 10 platforms
• 📊 Stats tab — your output at a glance

🎬 Demo

*Made end-to-end by ShortSmash — no hand-picked clips, no human editing. LLM script, neural voice, CLIP-verified stock footage, karaoke captions, ducked music bed.*

"How Goats Made You Addicted to Coffee" — 49s, 1080×1920, made from one prompt. Click to play:

Same video, re-voiced — native neural voices, same verified footage:

*Español · اردو · Roman Urdu · 中文 — the Mandarin version uses a native voice (not translationese) with Hanzi karaoke captions. One project, every audience.*

More demos — each made from a single prompt, click to play:

Why you can't tickle yourself — 11s, 1080×1920

The deepest place on Earth — 14s, 1080×1920

Why time flies faster as you get older — 11s, 1080×1920

📤 Or skip the AI — post your own video

Upload any video, get a pre-flight check (resolution, duration, per-platform limits), an AI-written title + hashtags, then post now or schedule it ("Friday 6pm") to 10 platforms. In the Studio's Publish tab, or one command: shortsmash post myvideo.mp4 --caption

🖥️ ShortSmash Studio

shortsmash serve — generate, browse trends, publish, and track stats from the browser:

Why ShortSmash

Most text-to-video tools assemble clips blindly — whatever the stock API returns first goes in. That's why automated shorts have the infamous "stock-footage look": a burger in a cooking tutorial, a parrot in a recipe video.

ShortSmash is the first open-source pipeline with eyes:

• Two-stage vision gate. Every candidate clip's middle frame is checked by a local CLIP model (offline, no quota, no cost) against 15 confuser categories — burgers, birds, houseplants, empty pans, citrus fruit — then by Gemini vision when quota allows. Wrong footage never ships.
• Audible gate. Renders with dead or missing audio fail instead of shipping.
• Never ships silent. A CC-licensed music library auto-downloads on first run; music is ducked under narration with a real FFmpeg sidechain.

The pipeline

prompt ──▶ script ──▶ voice ──▶ visuals ──▶ captions ──▶ render ──▶ publish

NVIDIA Edge-TTS Pixabay faster- FFmpeg official Gemini Piper Veo 3.1 whisper 1080x1920 APIs only Grok (offline Kling karaoke +faststart OpenAI- neural) Agnes compat 50 langs CLIP-verified

• Script — NVIDIA NIM (free) → Gemini (free) → Grok → any OpenAI-compatible endpoint (OpenRouter, DeepSeek, local Ollama) → offline template. Writes a scroll-stopping script *and* a strict JSON edit plan, validated by pydantic. Second-person, punchy, 3–4s scenes.
• Voice — Edge-TTS → Piper offline neural → gTTS fallback chain. 50 dub languages with native voices, Urdu/Hindi/Arabic included.
• Visuals — free Pixabay footage *or* AI generation: Google Veo 3.1 (best quality, native audio), Kling AI (top-tier motion), Agnes AI (free tier) — every clip verified by the vision gate regardless of source. Prefer your own? Set visual_provider: custom with any command (ComfyUI, Replicate, MiniMax…).
• Captions — faster-whisper word timestamps → burned-in karaoke captions.
• Render — FFmpeg: 9:16 smart reframe, transitions, hook + CTA overlays, ducked music, H.264/AAC faststart.
• Publish — YouTube, TikTok, Instagram, Threads, Bluesky, LinkedIn, Pinterest, X. Official APIs only — no Selenium, no scraping, nothing that gets accounts banned.
• Dub — shortsmash dub --lang ur re-voices any video in 50 languages — Urdu, Hindi, Arabic, Spanish, French, German, Chinese, Portuguese and 42 more — plus ur-roman / hi-roman for Roman-script captions ("Tum kaise ho?") with the native neural voice. Urdu captions render in proper Nastaliq calligraphy, right-to-left, with the karaoke highlight tracking the spoken word. Same verified clips, new native narration.
• Analytics — shortsmash analytics pulls real views/likes/comments from YouTube's API. Numbers from the platform, not vanity metrics.
• Trends — shortsmash trends harvests Google News + Hacker News (free, no keys). shortsmash batch turns the top trends into videos unattended.
• Schedule — natural language ("Friday 6pm"), durable SQLite queue, retrying worker.
• Engage — pulls recent comments, drafts warm replies, you approve each one with a tap. Nothing posts itself.
• MCP server — python -m shortsmash.mcp_server exposes generate_video, get_trends, list_projects, project_status to Claude Code, Cursor, and any MCP client.
• Resume — shortsmash resume picks up failed projects from the first incomplete stage instead of redoing everything.

How ShortSmash compares

*MoneyPrinterTurbo (129K★) is the most mature; OpenPost is the best manual workspace. ShortSmash is the most engineered AI-native pipeline — the only one that checks its own work.*

What ShortSmash will never do

No fake likes, comments, follows, shares, or stars. No botting, no "human-like" evasion of platform detection, no scraping or unofficial endpoints. That violates every platform's rules and gets accounts banned — and it doesn't work for real growth anyway. This boundary is architectural, not a policy note: the capability simply isn't in the codebase.

Quickstart — one command

**Windows (