Moritz Kremb ships FrameJam Playbooks for in-chat video editing

Moritz Kremb released FrameJam Playbooks, a video editing tool that lives inside the Claude or ChatGPT app rather than in a timeline editor. The described loop is four steps. You pick a playbook, drop in an idea or your footage, the agent makes the video, and you click anywhere on the result to leave feedback that the agent then fixes. Kremb calls it agent-native video editing and says it is free and open source.

Moritz Kremb

@moritzkremb

X

Click anywhere on the video to leave feedback. Your agent fixes them.

How you install it

Asked by George Ohan how a normie would use it, Kremb pointed at install commands on the website and said you paste them into the ChatGPT desktop or Claude Code desktop apps. He added that it all runs locally on your computer right now, which is the one deployment fact the posts commit to. There is no hosted version mentioned, no pricing page referenced, and no account step in the description.

Moritz Kremb

@moritzkremb

X

it all runs locally on your computer right now

What the posts do not say

This is a workflow announcement and not a measurement. The launch thread gives no render latency, no cost per video, no model name, and no statement of which decisions the agent is making between the footage going in and the cut coming out. Click-anywhere feedback implies something maps a point on a frame to an edit operation, but the posts do not say what performs that mapping or how often it gets it right. There is no eval, no before and after comparison against a conventional editor, and no failure case.

That matters more than usual here, because the hard part of agent video editing is the part nobody showed. Trimming, ordering and matching a cut to a stated intent are all judgement calls, and judgement calls are where cost and error rate accumulate. A demo that looks clean on one piece of footage tells you the interface works. It does not tell you what happens on an hour of unlabeled raw video, or what the loop costs when a user leaves twelve pieces of feedback on the same cut.

For people building on it, treat this as an interface pattern to borrow rather than a result to cite. The in-app install and local execution are the genuinely interesting bits, since they skip the upload and the render farm entirely, and local execution means your footage never leaves the machine. Everything else is unmeasured. If you try it, the numbers worth capturing on the first run are wall clock time per edit, how many feedback rounds a usable cut takes, and what the agent does when the instruction is ambiguous. Those are the figures the launch leaves for you to find.