Motion Skills
← Learn

The free, open-source motion video stack for coding agents: which tool does which step (2026)

Free, mostly open-source tools cover every step of a motion video made with Claude Code, Codex or Cursor: Remotion or Three.js for scenes, Manim for maths, CosyVoice for voice, WhisperX for captions, librosa for beats, FFmpeg to encode. None of them decides the story, the pacing or the facts.

Each project's facts are from its own README, licence file or site, checked on November 3, 2026. We make Motion Skills, so weigh what we say about it accordingly.

Lists of free "skills" for an agent motion studio are mostly libraries and apps, not agent skills: your agent can write code against them, but nothing in them says what makes a good film. Here they are by pipeline step.

At a glance

Step Tool What it does Licence What it leaves to you
Scene & motion Remotion Video as React components, rendered to a file Own licence; Company License for 4+ people Story, timing, look, sound
Scene & motion Remotion agent skills Teach an agent to write Remotion code No licence file; used with Remotion The same as Remotion
Scene & motion Anime.js Animates CSS, SVG, DOM attributes and JS objects MIT Turning a live page into video
Scene & motion Motion Animation library for React, JavaScript and Vue MIT (Motion+ is paid) Turning a live page into video
3D Three.js 3D in the browser, WebGL and WebGPU renderers MIT Lighting, camera, frame capture
Light three-gpu-pathtracer Path-traced light for Three.js MIT A WebGPU browser, render time
Physics Jolt Physics (JoltPhysics.js) Rigid-body physics, with a JS port MIT Wiring it to your scene
Effects pmndrs/postprocessing Bloom and other full-screen effects for Three.js Zlib Taste: when an effect helps
Effects three.quarks Particles and VFX for Three.js MIT Wiring, timing
Tuning Tweakpane A small panel for adjusting values live MIT It isn't part of the render
3D app MCP for Blender (formerly blender-mcp) Lets an agent drive Blender over MCP MIT; Blender is GPL Reviewing the code it runs
2D vector lottie-web Plays After Effects animations exported as JSON MIT Authoring, done in After Effects
Math Manim (Community) Python engine for maths explainer videos MIT The explanation, Python setup
Images & clips ComfyUI Runs image, video, 3D and audio models GPL-3.0 A GPU, each model's licence
Voice CosyVoice Multilingual text-to-speech and voice cloning Apache 2.0 (code and v3 weights) The script, rights to any voice
Captions WhisperX Transcription with word-level timestamps BSD-2-Clause Proofreading, caption styling
Beat sync librosa Audio analysis, including a beat tracker ISC Mapping beats to cuts
Capture OBS Studio Screen and camera capture, recording GPL-2.0 or later The take itself
Cut LosslessCut Trims and joins video without re-encoding GPL-2.0 Titles, effects, mixing
Encode FFmpeg Converts, encodes and muxes audio and video LGPL 2.1+, GPL with optional parts The exact commands

Scene and motion

Remotion writes video as React components and renders it on your machine, a server or AWS Lambda. Its licence is its own, not an open-source one: free for individuals, companies with up to three employees, non-profits and evaluation, "even commercial"; larger companies need a Company License. The licence file says it will "slightly change" in Remotion 5.0.

Remotion's agent skills (npx skills add remotion-dev/skills) teach Claude Code, Codex or Cursor the Remotion API: /remotion-create, /remotion-markup, /remotion-render, /remotion-captions, /remotion-maps and more. They make the code correct, not the film; see Remotion agent skills vs Motion Skills.

Anime.js and Motion (formerly Framer Motion) are animation libraries for live web pages, both MIT. Motion ships a free, MIT-licensed /motion agent skill (npx motion-ai); its Motion+ extras are a paid licence. Neither README describes exporting video: something has to step through the page frame by frame and encode it, or you record the screen.

3D, light, physics and effects

Three.js is the base: a general-purpose 3D library with WebGL and WebGPU renderers. three-gpu-pathtracer adds path-traced, physically based light; it needs WebGPU and builds each image up one sample at a time, so a clean frame takes many passes. Jolt Physics, used in Horizon Forbidden West, simulates rigid bodies; JoltPhysics.js brings it to JavaScript. Its simulation is deterministic, so a re-rendered scene falls the same way each time. postprocessing (Zlib) adds full-screen effects such as bloom, three.quarks adds particles, and Tweakpane gives you sliders to tune values while you preview.

MCP for Blender (renamed from blender-mcp) is a community plugin, not made by Blender, that lets Claude Code, Codex, Cursor and other MCP clients work in Blender. Its README warns that by default the agent "can run any Python code in Blender"; BLENDER_MCP_SAFE_MODE=1 blocks risky scripts. It sends a minimal anonymous usage record unless you set DISABLE_TELEMETRY=true. Blender is GPL, and its licence page says the images and films you make with it are yours.

2D vector and math explainers

lottie-web plays animations made in Adobe After Effects and exported as JSON with Bodymovin. The player is MIT; the authoring tool is After Effects, which isn't free. Manim, the community edition of the engine behind 3Blue1Brown's videos, renders maths explainers from Python scenes straight to video (MIT). The explanation is still yours to write.

Images and generated clips

ComfyUI (GPL-3.0) runs image, video, 3D and audio models in a node graph, locally: open models such as Wan, LTX-Video and HunyuanVideo, and closed ones through "Partner nodes". You need a capable GPU, and each model has its own licence, separate from ComfyUI's.

Voice, captions and beats

CosyVoice is a text-to-speech system covering nine languages and more than 18 Chinese dialects, with zero-shot voice cloning. The code is Apache 2.0, and so are the Fun-CosyVoice3-0.5B weights on Hugging Face. Clone only a voice you have the right to use.

WhisperX transcribes with word-level timestamps (wav2vec2 alignment) and writes .srt files. It runs on CPU, including on a Mac, or on an NVIDIA GPU. Speaker labels need a Hugging Face token and accepting the pyannote model's user agreement. Proofread names and numbers.

librosa (ISC) analyses audio; librosa.beat.beat_track is a dynamic-programming beat tracker. Which cuts land on those beats is your call.

Capture, cut and encode

OBS Studio (GPL-2.0 or later) captures, composites, records and streams; use it for screen demos and talking-head takes. LosslessCut (GPL-2.0) trims and joins files without re-encoding, using FFmpeg; it's free from GitHub, and the app-store versions support its developer. It cuts; it doesn't add titles or effects. FFmpeg does the encoding under nearly everything above. It's LGPL 2.1 or later, and GPL if built with its optional GPL parts.

What the stack leaves to you

Together these tools render, voice, caption and encode. You still decide what the film says and how long each beat holds, design the sound, check every number and name on screen against a source, and set up Node, Python, a GPU and a frame-by-frame render. That is most of the work.

Where Motion Skills fits

We make Motion Skills. A full-film skill is an agent skill that runs that pipeline for one kind of film (a launch film, an explainer, a data video, a short) on our engine, motion-skills, which is MIT. The skill carries the story, pacing and checks; the engine does the rest:

  • Scenes and motion: a film is HTML, CSS and GSAP (free under GSAP's own "no charge" licence), every frame a function of time, rendered frame by frame in the Chrome you already have. 3D shots use Three.js on the film's clock.
  • Music: an original score composed for the film, with scene cuts on the downbeat. mo beats finds the beats in a track you bring; mo music find looks for public-domain (CC0) music.
  • Voice: Kokoro voices or macOS voices on your computer, or ElevenLabs and HeyGen with your own API key.
  • Captions and checks: mo voice writes an .srt and, with --check, has whisper listen back to every sentence. mo check finds cut-off or covered text, mo audit matches every line on screen to sources.md, and mo qa scans the render.
  • Export: FFmpeg encodes MP4, MOV with alpha, WebM or GIF; a CapCut or 剪映 draft and a slide deck are options too.

It doesn't replace Blender or ComfyUI, and it doesn't make avatars. The Motion Director skill is free with a Google sign-in and makes 10–20 second films; full-film skills are $4.90 each or come with a plan (pricing, all skills). For other tools, including browser apps, see Claude Motion alternatives.

Questions

Is Remotion free?

For individuals, for-profit companies with up to three employees, non-profits and evaluation, yes, including commercial use. Companies of four or more need a Company License, sold on remotion.pro. Its licence is its own, not an open-source licence.

What free tools make motion graphics with Claude Code?

For code your agent writes: Remotion with its agent skills, Three.js, Anime.js, Motion and Manim. For the rest of the film: CosyVoice for voice, WhisperX for captions, librosa for beats, FFmpeg to encode. A free skill that runs the whole film is Motion Skills' Motion Director.

Can Claude make 3D animation?

Yes, through code or an app. Your agent can write a Three.js scene, with Jolt for physics and three-gpu-pathtracer for light, or drive Blender through MCP for Blender. Either way, check the camera, light and render setup.

Can I use these tools commercially?

The MIT, Zlib, ISC, BSD and Apache licences here allow it. The GPL and LGPL tools (Blender, ComfyUI, OBS, LosslessCut, FFmpeg) are free to use; their terms matter when you distribute the software. Remotion needs a Company License for companies of four or more.

Sources

This guide is independent. We're not affiliated with any project named here. Names are trademarks of their owners. Checked on November 3, 2026.