Commentary#Video generation#Video editing#Motion graphics
Top 10 Video Agent Skills Compared: Stars Belong to the Repo, So Pick by Output
We read the ten most-starred video skills. Stars belong to the whole repo, so pick by what the skill outputs. Includes a picker and the real limits of each.
On 2026-10-08 the ten most-starred video agent skills on GitHub add up to about 168,000 stars, but those stars sit on the repository, not on the skill you would install. The leader, OpenMontage, has 65.0k stars and no SKILL.md of its own. We published one skill from it, synthetic-screen-recording, and running npx skills add on the whole repo pulls about 91 folders, most of them third-party reference material.
We read the SKILL.md and README of all ten and published each as an entry on this site. This piece skips the per-entry detail and answers one question: which of the ten should you pick?
Key points
- The ten skills produce roughly five kinds of output: a rendered MP4, a re-runnable video project, a deletion list, Lottie or SVG motion, and a conversational editing session. Decide the output first, then look at stars.
- Most need something beyond the skill itself: a cloud ASR service, a third-party API key, a local app or cloud GPUs. Installing the skill does not make it usable.
- Licenses differ. The OpenMontage repo is AGPL-3.0, hypit’s modified Apache 2.0 bans hosted services and resale, and Remotion sells a separate license to larger for-profit companies. Read them before commercial use.
The top ten at a glance
All figures were checked on 2026-10-08. Stars come from the GitHub API; limits come from the SKILL.md and README files we read.
| Rank | Repo | Stars | What the agent produces | Biggest limit |
|---|---|---|---|---|
| 1 | OpenMontage | 65.0k | Terminal demos composed in Remotion | One skill depends on components in the repo; repo is AGPL-3.0 |
| 2 | hyperframes | 58.6k | MP4 rendered from HTML | Needs Node 22+ and FFmpeg; workflows download after install |
| 3 | hypit | 20.0k | An editable, re-runnable video project | Model costs hit your accounts; license bans hosted resale |
| 4 | lottie | 5.5k | Lottie JSON | Verified only in Skottie; last push was 25 July |
| 5 | skills (Remotion) | 4.9k | Videos written frame by frame in React | Repo has no LICENSE; larger companies buy a Remotion license |
| 6 | FireRed-OpenStoryline | 3.5k | MP4 from conversational editing | You install the app first and fill six model fields |
| 7 | chengfeng-videocut-skills | 3.0k | A deletion list | Codex plugin; needs cloud ASR |
| 8 | narrator-ai-cli-skill | 3.0k | Movie commentary video | Key from the maintainers; billed in points |
| 9 | pixel2motion | 2.4k | Smooth SVG and an entrance animation | Needs local Python, Chrome and Playwright |
| 10 | claude-code-video-toolkit | 2.2k | A script-to-MP4 pipeline | AI steps run on your own cloud GPUs |
Rendering video from code
The four skills in this group share one idea: the agent writes a program, the program renders each frame, so the result is repeatable and editable.
hyperframes is HeyGen’s entry skill and suits people who have not chosen a tool. It checks project state, interviews you into a BRIEF.md, picks one of 10 workflows from a routing table, then renders MP4 from HTML with data-* timing. The interview is heavy for a 10-second clip, and npx skills add installs only the entry, so the workflows download later through the CLI.
remotion-best-practices is Remotion’s official router skill, one of 12 skills in the repo. It sets three rules: open Studio before building, render only when asked, and animate with useCurrentFrame and interpolate instead of CSS animation. If your project is already React, choose it.
claude-code-video-toolkit puts skills, 14 slash commands, templates and a resumable project.json in one repo so Claude Code can take a video from script to MP4. Two cautions: installing only --skill remotion gets you the Remotion extension, not the pipeline, and the README’s own figures of about $0.01 per voiceover and $0.23 per AI clip run on your Modal or RunPod account.
synthetic-screen-recording comes from OpenMontage and solves a narrow problem: install and CLI demos are composed with TerminalScene instead of recorded, so re-renders are pixel-identical and typing follows the narration. For the project as a whole, see our OpenMontage entry.
Reworking existing footage
These four handle material someone else shot, or your own recorded talking head.
hypit hands back a project instead of an MP4: script, captions, B-roll and effects anchored to words, all editable and re-runnable. The README’s worked examples cost about $1.07 to $1.15 each. The license is a modified Apache 2.0: work for yourself and clients is fine, while hosted multi-tenant services, resale and removing the name from the CLI are not.
chengfeng-cut is the most restrained of the ten. It transcribes word by word, scans in five passes for repeats and slips, and produces only a human-reviewed deletion ledger, never touching the video. It is a Codex plugin installed with node bin/install.cjs; we did not test it in Claude Code, and the entry says so.
openstoryline-use has the agent start a local OpenStoryline service and revise a cut over many rounds with one session_id. You must install the app first, six LLM and VLM fields are mandatory, and AI transitions bill a third-party video model.
narrator-ai-cli wraps the hosted NarratorAI service, from movie search and voice to script and render, confirming at every step. It needs a NARRATOR_APP_KEY from the maintainers, bills in points, sends every request to the hosted service, and its SKILL.md says nothing about licensing for the film material.
Motion and logos
pixel2motion fits a PNG logo to smooth SVG, then animates it part by part into replayable HTML. A hard smoothness gate replaces an IoU target, so you get vectors you can animate per part instead of hundreds of jagged nodes. Every sample is a geometric logo; complex gradients are not shown.
text-to-lottie makes the agent check frame 0, the midpoint and the op-1 frame in a bundled Skia Skottie player before it counts the work as done, and it defaults to restrained design. The output is vector JSON for web and apps, not MP4. Skottie supports a subset of Lottie, which caps what it can verify.
How to choose
Ask three questions.
What output do you need? A finished video points to the rendering group, someone else’s footage to the reworking group, product motion to the motion group. This matters far more than the star count.
Where does the money go? hypit and the toolkit bill your own model accounts, narrator-ai-cli bills the maintainers’ points, chengfeng-cut needs a cloud ASR service. Read that column before installing.
Will it install into your agent? hyperframes and pixel2motion use npx skills add; chengfeng-cut is a Codex plugin; openstoryline-use needs the app first.
Our picks by situation: if you have not chosen a tool, start with hyperframes. If the project is React, remotion-best-practices. For a talking-head recording, chengfeng-cut to get the deletion list before cutting. For logo motion, pixel2motion for SVG and HTML, text-to-lottie for JSON in a web page or app.
Limits of this ranking
Stars measure repo-level attention, so rank is not skill quality. We read SKILL.md and README files and did not run every workflow end to end; each entry marks what is the README’s own claim and what we did not verify. Several projects move fast, and hyperframes was still receiving pushes on the day we checked, so look at the repo before you commit.