Image: Mortiflix
Why it mattersA marketer or founder who wants a short video gets a review step at every stage, so the final cut answers the notes they pinned at each one, instead of a single prompt returning one random result.
A marketer who writes a prompt and asks an AI to make a video gets whatever the model felt like making. Mortiflix, out this week, runs the job the way a motion design studio does it: a brief, a script, style frames, a transition board, an animatic, a final cut, and a review gate between each stage. The project is on GitHub as GTKottman/mortiflix-oss under AGPL-3.0 and sits at 349 stars in the two days since its public release.
The GitHub page describes Mortiflix as "a motion design studio on your own machine. Claude makes the video step by step. You approve every stage". The pipeline runs seven stages, from brief to delivered, and the session running the current stage stops at the end of it and waits for the reviewer. A note pinned on an exact frame or an exact moment of a video is answered in the next version, one note at a time, with the diff shown back.
Sessions that stop at the gates
The project README states a design choice that explains how a run that waits for a human can still cost nothing: a Claude session runs until it has something to show, writes a handoff and stops. When the reviewer answers, a fresh session reads the handoff and picks up from there. The project's words: "Waiting on you costs nothing, for hours or for days".
The rules a session must follow are written in code. Mortiflix refuses a submission that does not report every error check for its stage, or that does not answer each note left on the previous version. The reviewer's copy of the version is moved out of the session's reach after the gate, so a later session cannot edit it.
It learns the reviewer's taste
Point out a real mistake, and the session can propose a new check, such as "text never touches the frame edge". Approve it once and every future video the studio produces runs that check. The learned preferences live in a TASTE.md file the project carries across days and jobs, so a reviewer builds up a studio rather than retraining the model.
Pipelines live as folders in the project. A pipeline is a pipeline.json listing the steps, how each one is reviewed, and the error checks that must pass; a PIPELINE.md for the craft notes; and the skills Claude Code loads to do the work. The project says "anyone can write one, that's the point of open-sourcing it".
Install and limits
Node 20 or newer and FFmpeg are required. Claude Code with a plan or an Anthropic API key is required for real work. Mortiflix states that it is "tested mostly on Linux so far" and that a Windows version is due by 20 October 2026. macOS is untested.
Two ways to run it: mortiflix new explainer in the terminal, with mortiflix run to progress and mortiflix review to answer, or mortiflix serve for a browser at 127.0.0.1:4646 that shows the review room and a live log of what Claude is doing. The project includes a mortiflix demo that walks the full pipeline with placeholder work, so a reader can try it without a Claude Code login or an API key.
Source
- GTKottman/mortiflix-oss on GitHub (the project README)
- How to use Mortiflix on YouTube
This item was written by an AI system from the linked source. Reveneau is responsible for what it publishes.
Get AI News in your inbox
New developer tools, model and agent releases, and how teams are actually using them to release software. Short, and only when there is something worth reading.