Lythara
A video editing agent designed to understand multiple clips, personal taste, and the idea behind a video.
Lythara began because I wanted an editing agent that could understand my footage, my taste, and the idea I was trying to communicate.
The situation.
I was getting frustrated with manual video editors. They demanded a lot of repetitive cutting and arranging, and I was not interested in paying high subscription prices for plans that did not provide enough value for my own workflow.
I wanted something more personal than a generic automatic editor. The idea was an AI video editing agent that could read multiple clips, understand what each one contributed, learn the kind of pacing and choices I preferred, and place the footage according to the idea I was trying to portray.
What I did.
I set the independent product direction, interface design, architecture, and source-available implementation. I chose a local desktop workflow so original clips could stay on the creator's machine while the agent analysed the available media and prepared an editable first pass.
I designed the agent around intention rather than a fixed template. A creator should be able to describe what the video is trying to say, give the system several clips to reason across, and receive an arrangement shaped by that brief and their taste. The timeline remains available for checking and changing every decision.
How the work unfolded.
Reason across multiple clips
Designed the workflow to examine a collection of footage as a whole so clip selection and ordering could serve the video's idea rather than follow upload order.
Direct the edit through intent and taste
Used a natural-language brief for mood, pacing, grade, story, and personal preferences instead of forcing every project through the same preset.
Keep every decision editable
Provided timeline actions such as split, reorder, duplicate, delete, trim, scrub, and play so the generated first pass remained a proposal rather than a locked result.
Bound the project's collaboration model
Published source for inspection and forks while documenting that external pull requests and code contributions are not accepted.
What happened.
Lythara's public repository contains the phase-zero technical spike and first usable local editing loop. It can analyse local footage, work from a creative brief, and produce an editable first pass while leaving the creator in control of the timeline.
The spike turned the original idea into a concrete workflow: give the agent several clips and an explanation of the video you want to make, let it reason across the footage, then inspect and change every choice on a normal editing timeline.
What I learned.
Taste is part of the specification
A useful personal editor needs to learn what the creator prefers, not simply apply a generic definition of a polished video.
The agent needs the whole set of footage
Choosing a clip well requires context from the other clips and from the idea the finished video is meant to communicate.
Local workflow is a trust and cost decision
Keeping source media local gives creators clearer control over their files and avoids making another expensive cloud subscription the centre of the product.
Start with the workflow before the platform
Phase zero keeps the focus on whether a directed, editable first pass is useful before expanding into a general-purpose editor.
Related links.
Repositories, demos, event pages, posts, and other places connected to this project.