Gitframes ships professional video editing as a WebGPU npm package for AI agents

A project called Gitframes has landed on the Hacker News front page with an ambitious pitch: fold the work normally spread across Photoshop, After Effects and Blender into one npm package that AI coding agents can drive with TypeScript, then render the result straight to an MP4 on the GPU.

One package, three toolchains

According to the project's README on GitHub, published under the gatewai-dev organisation, Gitframes aims to cover three categories of desktop tool in a single library: 2D compositing and visual effects, motion design with typography and keyframing in the After Effects tradition, and 3D scenes with cameras and models. The intended workflow is that an agent writes a TypeScript composition, samples individual frames to check the result, and renders the finished video, with no multi-gigabyte creative suite to install or license.

The feature list is long. Typography uses the Slug algorithm, evaluating glyph outlines analytically per pixel in WGSL shaders, which the developers say keeps text crisp at any resolution and enables After Effects-style range selectors and animator behaviors. Post-processing offers more than fifty GPU shader nodes, including curves, levels, selective color, 3D LUTs, halftone, film grain, unsharp mask, mesh warp and screen-space relighting. The 3D side is a unified scene graph with camera rigs and loaders for OBJ, FBX, glTF and several other mesh formats, along with PBR glass and SSAO. Audio is handled by native DSP with multi-track mixing, procedurally generated transition sounds, and signals derived from audio that can drive visuals. There are also animated chart layers and on-device vision features, such as object detection, instance segmentation, pose estimation and person mattes, running on Apache-2.0 licensed ONNX models.

Rendering without a browser

The architectural argument is the sharpest part of the pitch. Much of today's programmatic video tooling, Remotion most prominently, renders by driving a headless Chromium instance and capturing screenshots. Gitframes skips that step: it talks directly to native GPU devices through Dawn, the WebGPU implementation that reaches Metal and Vulkan backends, inside Node.js or a WebGPU-capable browser, and it encodes video in hardware through the napi-rs WebCodecs bindings. The project also promises deterministic, frame-exact timing, with an absolute frame clock and frame-aligned audio grids rather than floating timers.

The README includes a comparison against Remotion, which is built on React and Chromium, and Hyperframes, which uses Canvas2D and SVG. The project's own figures put Gitframes at 60 to more than 120 frames per second, against a claimed 5 to 20 for Remotion and 20 to 40 for Hyperframes. Memory use is listed at roughly 200 to 400 MB per render versus 1.5 to 4 GB for a Chromium-based worker, with a container image around 500 MB instead of 2 to 3 GB. These are the developers' numbers rather than independent benchmarks, and the table does not weigh ecosystem maturity or community size, where Remotion currently has a clear lead.

Built for agents and pipelines

Agent integration is treated as a first-class concern. Claude Code users can add the Gitframes skills with a single plugin install command, and other coding agents, including Codex, Cursor, Gemini CLI and Copilot, can pick up the same skills via a one-line npx command. For verification, the library provides FrameGrid contact sheets, single-frame snapshots and pixel-comparison assertions, so an agent can check its own output much the way it runs tests against ordinary code. The modest memory footprint is pitched as a good fit for serverless GPU clusters, with the README naming AWS G4 and G5 instances, Modal, RunPod and Kubernetes.

Caveats

The README is blunt that this is beta software: APIs may change between releases and some features are incomplete or unstable. Teams building on it today should pin versions and expect churn.

Why it matters

If the performance claims hold up, Gitframes reframes programmatic video from browser automation to conventional software engineering: deterministic output, a small memory footprint, frames that can be asserted against in tests, and rendering that fits CI pipelines and serverless GPU fleets. For AI coding agents, it turns video into just another artifact they can generate, inspect and verify in a feedback loop. It is also a signal of a broader shift: as WebGPU support matures in Node.js, GPU-native rendering is moving out of game engines and desktop suites and into plain npm packages. The beta label means production adoption is premature, but the architectural bet, frame-exact video without a browser in the loop, is the part worth watching.