If you have ever wanted a VTuber-style avatar but bounced off the weeks of manual rigging that tools like Live2D demand, Mesh Avatar Studio tries a different deal: hand your illustration to a coding agent, let it build the rig, then fix the details by dragging dots in a local editor. The result blinks, talks, turns its head, breathes, and sways its hair, and it is MIT-licensed open source.
Key Takeaways
Mesh Avatar Studio (github.com/shinshin86/mesh-avatar-studio) turns one illustration into an animated 2D mesh avatar: blinking, talking, head turns, breathing, hair sway.
A coding agent (Claude Code or Codex) does the initial rigging by following the repo's agent guide; you refine it in a local drag-to-fix editor.
It ships with real-time face tracking (MediaPipe), OBS streaming support, a portable `.mavatar` format, and an embeddable `mesh-avatar` npm runtime.
MIT licensed, with 795 stars and 90 forks as of October 18, 2026, only two weeks after its creation on October 4.
What Is Mesh Avatar Studio?
Mesh Avatar Studio is an open-source project by Yuki Shindo (known on GitHub as shinshin86) that converts a single illustration, ideally a front-facing, head-and-shoulders PNG with a transparent background, into a rigged 2D avatar. The avatar is driven by mesh deformation in WebGL: it blinks, lip-syncs speech, turns and tilts its head, breathes idly, and sways its hair.
The repo is MIT licensed and moving fast: 795 stars, 90 forks, 120 commits, and 4 releases since its creation on October 4, 2026, with the last push on October 17. Its GitHub topics read like a mission statement: 2d-animation, avatar, claude-code, codex, mesh-deformation, vtuber, webgl. A 7-minute video walkthrough shows the full flow from a PNG to a tracked avatar.
How It Works: Your Coding Agent Does the Rigging
Making an avatar from your own illustration takes four steps, and the division of labor is the interesting part. Steps 1 and 4 are done by a coding agent; steps 2 and 3 happen in the editor.
1. Ask your agent to create the avatar. Open Claude Code or Codex in the repository folder, hand it your illustration, and tell it to follow `docs/agent-guide.md`. The agent checks how accurately it reads coordinates, places the rig on zoomed grids, cuts the image into layers, and reviews the avatar in fixed poses. The recommended setups are Claude Code with Claude Opus 5.5 or Codex with GPT-6.1 Sol, per the README.
2. Open the project and check it. Start the local editor (`npm run dev`, a Vite dev server) and open your project. Use Pose test to turn and tilt the head and close the eyes, and Lip sync to check the vowel mouth shapes.
3. Fix what looks wrong by dragging. Select a body part on the left; its control-point dots are highlighted on the image. Drag them to fix the placement and watch the live preview follow. Changing outlines that affect layer cutting (eyes, hands, head, hair) triggers a "Save and rebuild layers" step that takes a few seconds.
4. Get drawn eyes and mouths. Without drawings, eyes and mouth move by mesh deformation alone. Drawn closed eyes and vowel mouths make blinking and talking look much more natural. Here the agent choice matters: Codex draws them with its built-in image generation and imports them automatically, while Claude Code cannot generate images, so it prepares the masks and prompts for you to hand to an image generator. An optional Expressions pack adds drawn sad and surprised eyes plus closed angry and sad mouths.
From Editor to Livestream
A rigged avatar is only half the story; the project wants it on stream.
Face tracking that stays local. The Live page uses your webcam with MediaPipe's Face Landmarker (vendored in the repo under the Apache 2.0 license) for head pose and expressions, plus microphone lip sync. The camera is optional: with only the microphone on, the avatar keeps its idle motion and your voice moves the mouth. The privacy design is deliberate: camera video and microphone audio never leave your machine; only numeric motion values reach the stream view.
OBS-ready. Copy the OBS URL from the Live page and add it as an OBS Browser Source (1080x1080 suggested). The background can be transparent or green for chroma keying. Expression hotkeys (1-8) and draggable-light lighting presets (Shift+1-8) switch live, even while OBS is in front.
Portable avatars. Export a single `.mavatar` file containing the rig, built layers, drawn expressions, and metadata (imports up to 128 MiB), and share it. The `mesh-avatar` npm runtime lets developers embed avatars in their own apps with plain WebGL and TypeScript types, no React dependency: `npm install mesh-avatar`, then `loadMeshAvatar(canvas, '/avatars/character.mavatar')`.
Zero-install player. If you just want to try it, the browser Player runs face tracking, lip sync, expressions, and lighting entirely in the browser. Open the sample or your own `.mavatar` file; nothing is uploaded.
Who Made It
The author is Yuki Shindo, who writes on his GitHub profile: "I love building things. I develop software for AITubers, and lately I've been really into vibe coding." He runs AITuber OnAir, a fully browser-based AI avatar streaming tool, and his open-source track record includes oh-my-logo (1.6k stars, ASCII-art terminal logos) and an aituber-onair toolkit (281 stars).
One licensing nuance worth knowing: the code is MIT, but the bundled sample character Miko belongs to the author's AITuber OnAir project. Her images ship under separate character usage guidelines: you may use and modify them as part of your own works, but not redistribute them standalone or as an asset collection.
How It Compares to Live2D
The obvious reference point is Live2D Cubism, widely described as the industry standard for 2D rigging and the tool behind nearly all professional 2D VTuber models.
Aspect | Mesh Avatar Studio | Live2D Cubism |
|---|---|---|
Price | Free (MIT open source) | Free tier (100 parameters); Pro indie about $100/year |
Rigging | Coding agent builds the first rig; drag-to-fix editor | Manual rigging with a steep learning curve; most streamers hire riggers |
License | MIT (code) | Proprietary |
Runtime | Embeddable `mesh-avatar` npm package (WebGL) | VTube Studio and proprietary SDKs |
Best for | Developers and tinkerers who want an avatar this weekend | Professional VTuber models |
The contrast the project is betting on: Live2D rewards skill and patience, while Mesh Avatar Studio tries to make the first rig something an AI agent does for you.
Things to Know Before You Try It
A few honest caveats, all from the README itself. You need Node.js 22.17+ and Git. The best input is a front-facing, head-and-shoulders PNG with a transparent background. Drawn eyes and mouths are easiest with Codex, since Claude Code cannot generate images. The sample character Miko is not MIT-licensed, so check her usage guidelines before redistributing. And while the project is only two weeks old, it is under very active development (120 commits, last push October 17, 2026), so expect rough edges and fast changes.
Frequently Asked Questions
Do I need to know how to rig?
No. That is the point. The coding agent builds the initial rig from your illustration by following the repo's agent guide; you only adjust control points by dragging in the editor.
Which coding agent should I use?
The README recommends Claude Code with Claude Opus 5.5 or Codex with GPT-6.1 Sol. Pick Codex if you want AI-drawn eyes and mouths, since Claude Code cannot generate images.
Can I use it for real VTuber streaming?
Yes. The Live page gives you an OBS Browser Source URL with a transparent or green background, face tracking, microphone lip sync, expression hotkeys, and lighting presets.
Is everything really free and open source?
The code is MIT licensed. The exception is the bundled sample character Miko, whose images follow separate character usage guidelines.
Do I need to install anything to try it?
No. The browser Player runs face tracking and lip sync entirely in your browser with nothing uploaded. Installing is only needed to create and edit avatars.
Conclusion
Mesh Avatar Studio is a small, fast-moving answer to a big annoyance: rigging a 2D avatar by hand takes skill most people do not have. By handing the first rig to a coding agent and keeping a human in the loop with a drag-to-fix editor, it compresses weeks of specialist work into an afternoon project. At 795 stars in its first two weeks, with an MIT license and an embeddable runtime, it is one of the more interesting open-source plays in the VTuber tooling space right now. If you have an illustration sitting in a folder, this might be the weekend it starts talking.