The MOSA Video Engine is a production company that lives in one codebase: a crew of twenty-five AI roles, forty-two templates, and a stack of model APIs, directed by one art director. Its first real test was Earthrove, a wildlife documentary channel. In thirty-five days it shipped six nine-minute films and six Shorts, with fourteen more Shorts finished and waiting. This is how it works, what it did, and what broke.
The first five Earthrove films, as their live thumbnails. Each is about nine minutes. All five were made by the engine.
A real wildlife film takes a researcher, a science editor, a writer, a narrator, a composer, a footage producer, a motion designer, an editor, a colorist, a sound mixer, a thumbnail designer and someone to publish it. Most "AI video" skips most of those jobs, and it shows.
I didn't want a faster way to make worse videos. I wanted the whole studio. So I built each of those jobs as a seat in a system: a written brief, a set of tools, and a quality gate it has to pass. Then I kept the calls that need taste for myself.
The engine started on July 14, 2026 as one folder. It has been the video department for everything I make ever since: client product films, a cinematic main title, broadcast spots, music videos, and the channel this page is about.
It's built on Remotion, which renders video from React. That means a film is a program: it can be versioned, reskinned and reused. Change one brand file and every template in the engine takes on a new palette.
No single AI model makes a film. The engine calls each one for what it does best, checks the price first, and logs where every asset came from.
4K stock video through the API, every download logged with its license and provenance. Image-to-video for illustrated title slates. Cost simulated before anything is spent.
Seedance 2.5 video with a strict JSON prompt dialect, Nano Banana stills, and Suno V6 scores at 12 credits a track.
Narration with word-level timestamps, so captions never drift. Every take is transcribed back and diffed against the script before it's used.
Free stock is always searched first, and archives like NPS, USFWS and Wikimedia Commons are searched before anything paid.
Diagrams and mechanisms drawn in three.js, WebGL and canvas. From the monarch film on, the explanatory animation costs nothing to make.
Unreal Motion Design for promos, Blender for grading, and After Effects driven through MCP: one 30-second ad was built 121 layers deep by script.
The rule that keeps it affordable is a ladder: code first, then archive, then free stock, then local 3D, then AI stills, then AI motion. The beaver film used 37 stock masters at zero credits; its 3D beaver cost 135 credits and its score 36. When an upscale was quoted at 36,297 credits, the engine took a rule from it: never upscale a whole sequence, render at the target size.
Five Ways to Live in the Ocean · 9:00 · Earthrove's best-retaining film so far (41.7% average viewed). Loads from YouTube only when you press play.
For scale: five days after launch, on September 13, the channel had 26 views and 1 subscriber.
84% of impressions came from YouTube recommending the videos to people who weren't looking for them. The Shorts carry the reach, with two thirds of all views. One of them, on the mantis shrimp's punch, is watched to 89% of its length on average. And this is a living-room channel: 60% of all watch time happens on a television.
Read honestly: 10.8% of views came from YouTube advertising, and 99.7% of the audience is new. The channel finds viewers; it doesn't bring them back yet. Source: YouTube Studio analytics for @earthrove, Sep 7 – Oct 11, 2026, read October 12.
The engine keeps a log of everything that went wrong, dated, with the fix. It reads that log before every job. These are the ones that taught it the most.
The research bench earns its seats too. It caught a 2014 paper that had been retracted before it reached the bear script, and a salmon figure that came from the wrong species.
Viewers drop at about two minutes, where the first chapter ends, and almost nobody comes back for a second film yet. So the next builds are structural: chapters that open on a new question, films commissioned as a series, and end screens that hand you the next one.
Packaging is already in test. Click-through was 5.4% on the first films and is 3.3% across the channel's lifetime, so the engine designed challenger thumbnails for all five films from one rule set (a real frame, four words at most, one highlight color, legible at 168 pixels) and started YouTube A/B tests on October 2.
Earthrove isn't the business. It's a stress test. If one art director can run a documentary channel this way, with research, voice, score, edit, QA and packaging, then your team can run its campaigns, product films and social pipeline this way too. It runs on your brand, your gates and your calls, with a system that remembers every mistake so you only make it once.
Production notes: Earthrove's narration uses ElevenLabs voices and its scores are generated with Suno. Illustrated sequences use AI image and animation tools and are labeled as illustration. The earliest channel setup and production work, September 7 to 12, ran with OpenAI Codex as the AI partner; the engine and everything after ran with MOSA, my AI studio partner. Metrics are from YouTube Studio as of October 11, 2026.