14 days, 470,000 lines of code: OpenMotion turns viral videos into editable templates
Tom Huang has open-sourced OpenMotion, which lets you recreate brand promo videos with a local Agent, export MP4s, and keep the HTML project. It writes shot order, text rhythm, and transitions into MOTION.md, with 300+ built-in templates and 152+ brand kits.
OpenMotion is now officially open source. Tom Huang announced the project on X, and its positioning is straightforward: break a reference video into shots, text rhythm, transitions, and sound design, then recreate it with a local Agent and an editable project. The project is on GitHub and licensed under Apache 2.0.
A common problem with video generation tools comes later, during revisions: once the video is done, wanting to swap the product, change one line of copy, or adjust the pacing by two seconds often means rerunning everything. OpenMotion's approach is to write the video as HTML, CSS, and JavaScript. The Agent breaks down the reference video, rebuilds shots, and fills in assets; Studio handles timeline editing; and finally you export an MP4 from the same project, with the source files kept locally.
Tom Huang says the team wrote 470,000 lines of code in 14 days, reverse-engineered product promos from 100+ leading companies and startups, and organized 300+ editable templates plus a motion-focused DESIGN.md, with 152+ Brandkits built in. The fidelity exceeds 95%. He also says you don't need SOTA models like Opus 5.5 or Fable 5.2; ordinary Agents can run it too. The templates cover product promos, brand films, data charts, music visuals, talking-head edits, and other categories. After swapping brand assets, copy, and product footage, you can keep the same motion language.
The core file of this system is called MOTION.md. It writes shot order, text rhythm, transitions, and sound design into a spec that an Agent can read, and you continue editing with brand assets and the project. The reference video leaves behind both the feel and executable production notes.
The integration follows the Agent workflow. OpenMotion supports 20+ Agents including Codex, Claude Code, DeepSeek Harness, and WorkBuddy, invoked through Skills, MCP, and plugins, and connects to 74 vendors. It also supports BYOK and integrates multimodal models such as Seedance 2.5 and ElevenLabs, so image generation, video generation, voiceover, and music can all run inside the project. After connecting via plugin or MCP, you can make requests in the original conversation or directly operate the editor to make changes. The README also provides a one-click install entry for ChatGPT Marketplace.
The project team calls it full-stack: the Agent ideates, generates assets, and arranges motion; you swap assets, adjust duration, and change pacing in the same project, then export. If you want to keep editing, hand the requirements back to the Agent. Files are kept locally, with cloud processing planned later.
The open-source part builds on community projects. OpenMotion thanks projects including HyperFrames, Hypit, rea-video, GSAP, FFmpeg, and OpenChatCut. HyperFrames provides the rendering foundation for converting HTML to video, Hypit provides the idea of using an Agent to clone viral videos, and rea-video contributes reference analysis cases. The project uses Apache 2.0, and the code is at https://github.com/nexu-io/openmotion .
In the comments, some users said there were so many templates they were spoiled for choice, while others suspected it was a thread ad. Tom Huang replied that this is a product he open-sourced himself and that people are welcome to try it. Project maintenance resources are tight, so anyone with issues can file an Issue or PR.
Charlie posted four movie-opening examples on X. Star Wars, The Matrix, F1, and Marvel each correspond to a remixable MOTION.md. For people making product launch films, this is more useful than one-click generation: it leaves behind a structure that can be further edited.
发布时间: 2026-11-17 23:36