Gitframes, an open-source toolkit for producing video through code, is positioning WebGPU as the rendering foundation for a workflow aimed at developers and AI coding agents. The project packages compositing, motion graphics, typography, keyframing and 3D scene tools into an npm package, allowing a composition to be described in TypeScript and rendered as an MP4.
The project remains in beta, and its maintainers warn that interfaces may change and some functions may be incomplete or unstable. That caveat matters for teams considering it for production use: Gitframes is presented as an actively developed engineering toolkit, not a finished replacement for established editing applications.
Its central technical choice is to render directly through GPU interfaces rather than automate a general-purpose browser. The repository says it can use Dawn, WebGPU, Metal or Vulkan in Node.js and modern browsers. By avoiding a DOM layout and screenshot-capture pipeline, the project seeks to make programmatic video behave more like a conventional software build, where source-controlled code defines each frame. Node.js 22 or newer is required.
Gitframes combines several types of visual work that are often divided between separate applications. Its layer system includes paths, boxes, text and charts, with chart geometry and scales calculated through D3. The repository also describes shader-based image and video effects, 3D cameras and models, and tooling for inspecting rendered frames. Example compositions are included as references for developers building their own scenes.
An optional vision package connects ONNX models to the same reactive signal system used by other elements in a composition. According to the project documentation, creating a vision runner does not immediately download a model or open a session. Model files are retrieved only when a task first runs, unless a developer explicitly preloads them. The listed models are pinned to specific revisions and checked with SHA-256 hashes after download.
The agent-focused layer is instructional. Gitframes distributes skills intended to teach tools such as Codex, Claude and other coding assistants how to create, render and inspect compositions. The project says these skills contain instructions rather than executable servers, hooks or launchers. The rendering engine itself is installed through npm, while the skills can be installed through supported agent ecosystems or copied into the relevant skills directory.
Gitframes is licensed under Apache-2.0, as are the vision models identified in its repository. The code-first approach could make video assets easier to review, reproduce and automate alongside other software, but the beta label means adopters should expect compatibility changes and validate outputs before relying on the toolkit in a production pipeline.
