lazyreel

LazyReel Plugin

An AI agent that doomscrolls for you.

Software has spent forty years being built for a human clicking around a screen. That era is ending. The agent is the customer now, and LazyReel is built for it: this is the beginning of agent-first marketing, where you hand the whole ideation phase to a system that scrapes and remembers far more than you ever could.

LazyReel is a free MCP server that already doomscrolled 21B+ views of short-form video (and app ads) so your agent doesn't start from zero. Plug it in and your agent knows what is breaking out in your niche: the hooks, the formats, and the real videos behind them, before it writes a single line. You delegate the whole ideation phase, what is worth making and why; it hands back the decoded research and the brief.

This is the taste layer, the brain. Generation tools like Higgsfield are excellent hands, they will render anything, but they do not tell you what to make. LazyReel decides what is worth making and hands over the exact brief; the gen MCP renders it. Same agent, same chat. You stop shipping slop because what you generate is grounded in what actually broke out.

It is the first layer of an agent-first marketing stack. Decode with evidence, not vibes.

What it gives your agent

6 tools (plus a status tool). Each one composes the underlying analysis, so a single call returns a full report, not a thin slice.

1. niche_report. The full read on what is working in a niche, from real decoded videos: the hook patterns and frameworks that over-index in breakouts (with lift and sample size), the words winners open with, real example links, the winning feature combinations, and the supply/demand gap nobody is filling. focus narrows it: format (how winners look), trends (cross-niche formulas), combos (the winning mixes), apps (the app-ad vertical).
2. study_videos. Real TikTok links to watch and study, best breakouts first (each with reach multiple, views, format, hook, emotion), plus a diagnosis of why the top ones out-reached their creator's following. Filter by niche/format/hook, or pass a free-text query.
3. teardown. Reverse-engineer a format. Pass a video (description, transcript, or URL) for its narrative + edit DNA and how to rebuild it; or pass a product to get the proven format, real examples, a shoot-ready brief, and a generation-prompt scaffold.
4. make_brief. Write the creative. Default returns a shoot-ready brief (framework, angle, hook bank, a Beat/VO/on-screen-text/B-roll table, CTA, shoot notes); mode=ideas for concepts, mode=hooks for a ranked hook bank.
5. breakout_laws. Why the same concept gets 1K vs 1M views: the validated first-3-seconds laws that separate breakouts from duds (concept-matched and same-creator pairs plus corpus lift; 83% within-creator on the cleanest blind test, see docs/methodology). Returns the laws, what over-indexes, matched proof, and an honest confound caveat.
6. kill_the_slop. Point it at weak copy. It names exactly what is wrong against the anti-slop bar, then rewrites it sharper using a named hook pattern.

And get_status, which says plainly what is live and what is not. No marketing in the tool output.

These 6 tools are the research layer. The 6 companion skills in skills/ are the make-it layer (deconstruct, the Seedance and Higgsfield directors, the video editor) that turn this research into the prompts and the finished cut. So the product is 6 tools + 6 skills.

What is inside

The skills are grounded in real DTC and UGC scripting knowledge, not vibes:

• 12 named script frameworks (PAS, the contrarian, the listicle, the founder story, and more)
• A 13-pattern hook taxonomy (POV, belief-challenging, direct call-out, size of claim, and the rest)
• 6 proven UGC angles with shot counts
• Awareness and sophistication models, so the angle matches where the buyer's head is at
• A hard anti-slop bar: banned filler words plus the voice rules that separate real creator copy from AI mush
• A library of analyzed video teardowns across the main DTC niches (beauty, skincare, supplements, fitness, food, tech, fashion, home, hair, pets), each tagged with its framework, hook pattern, signature device, and why it worked
• Real TikTok trending-hashtag data for 2022 through 2025 (MIT licensed) so niche reads can point at live cultural moments

The corpus lives in mcp/data and grows over time. See mcp/data/DATA.md for sources and licenses. The study_videos tool queries it directly.

How the tools point at real videos: the committed reference list is mcp/data/examples.json, 418 real videos with their public links and decoded labels (niche, hook pattern, format, views, views-per-follower). That file is what study_videos reads to hand your agent real videos to study. The raw scrape and the downloaded video files stay out of the repo. Only the public links and the derived analysis are committed.

The window, and why recency wins