ManimGX: a 3D animation engine for coding agents

Sina Atalay (verified account @sinaatalay) open-sourced ManimGX, a 3D animation engine built for coding agents, on October 11, 2026. It uses a Rust and wgpu core while targeting compatibility with Manim's Python API, and ships with an open repository and docs.

Image source: @sinaatalay via X

The release message is summed up in the author's own line: for 40 years we built GUIs for humans, and now we build APIs for agents. The launch post even suggests pasting a one-line prompt into a favorite coding agent to generate a 3D animation of the solar system and open it.

Rust + wgpu core targeting the Manim Python API

ManimGX is built on a Rust and wgpu core and is designed to stay compatible with Manim's Python API. In a follow-up post, the author thanks Manim creator Grant Sanderson and ManimCommunity, noting that ManimGX targets the ManimCommunity API.

The author also lists the projects it builds on: wgpu, Google's Filament, Typst, mitex, x264, FFmpeg, libopus, Pyodide, and Mesa. The repository and documentation are linked from the launch post.

Install path and author-claimed performance figures

Per the author's description, setup is a single line — pip install manimgx — supporting macOS, Windows, and Linux on Python 3.13+, with a 65–110 MB download and no system dependencies such as LaTeX or FFmpeg. A browser path is also mentioned: npm install manimgx, with Pyodide and WebGPU rendering the scene.

All performance figures are author claims and have not been independently verified. The author states that 36 seconds of 1080p60 video renders in 3.1 seconds from launch to finished MP4, and claims 90.6× over ManimCE, 11.5× over ManimGL, and 18.7× over Blender Workbench, citing a same-author follow-up post and a self-provided benchmarks link. This article treats those numbers as claims only.

Agent-oriented features: inspect, layout checks, narration

Several features target agent-driven workflows. A manimgx inspect CLI command produces storyboard PNGs an agent checks before rendering the video, and 2D layout checks flag cut-off content, overlapping or too-small text, by variable and line. Static typing support (TypedDict keyword arguments, typed .animate chains) is also advertised.

Voiceover support is built in through self.say(). For example, self.say("The derivative of [x squared] is [two x].", anim1, anim2) speaks a line and starts each animation at its bracketed words; any text-to-speech service can supply the voice, and rendering writes the captions, according to the author.

Developers who script math animations or explainer-video pipelines in code — and teams that want a coding agent to drive scene generation — may find this tool worth a look. Since the speed comparisons rest on author-provided benchmarks, measuring with your own scenes and hardware before adopting is the safe move.

Sources

• ManimGX GitHub repository: academa-labs/manimgx
• Sina Atalay launch post: Introducing ManimGX
• Author follow-up post (benchmark claims): render speed claims
• ManimGX docs: manimgx.academa.ai