Mesh Avatar Studio turns a single illustration into an animated talking 2D avatar through a Claude Code or Codex workflow

Making a hand-drawn character blink, speak and turn its head has required a rigging specialist and a licensed tool, which has kept the whole category closed to a one-person studio with one illustration. Mesh Avatar Studio is an MIT-licensed project from Yuki Shindo that moves the rigging job to a coding agent, with the artist stepping in afterwards to drag the dots that are wrong. The GitHub REST API reports the first public commit on 4 October 2026 and 119 stars one day later.

The workflow has four steps. In step one the artist opens Claude Code or Codex in the project folder and gives it a path to a PNG. The agent follows a procedure in docs/agent-guide.md: it checks how accurately it reads coordinates, places the rig on zoomed grids, cuts the image into layers and reviews the avatar in fixed poses. In steps two and three the artist runs a local editor, drags handles over the character's own art, uses pose and lip-sync previews to see what still looks wrong, and saves. Step four, which is optional, generates drawn eye closes and vowel mouths from Codex's own image generation, which the editor reloads automatically.

What the agent handles and what the person handles

The division of work is the useful detail. The agent reads the illustration, maps its coordinates, cuts it into layers and lays down the rig, which are the steps that an artist without rigging training usually cannot do. The artist drags the dots on top of their own picture, which is the step an agent cannot do as well as a person looking at a character they drew. A banner appears when a drag changes a value that affects how the image is cut, and the editor rebuilds the layers in a few seconds.

Shindo recommends Claude Opus 5.5 for Claude Code and GPT-6.1 Sol for Codex, named in a two-row table in the README. Only Codex can draw the optional closed-eye and mouth shapes, because Claude Code cannot generate images. In the Claude Code case the editor instead produces the masks and the prompts for the artist to hand to a separate image generator.

The sample and its licensing

The bundled example is Miko in a qipao, from the project AITuber OnAir. The code is MIT licensed, but the sample character is not: Shindo states that Miko's images are provided under the Miko Character Usage Guidelines, which allow use and modification inside a user's own works and do not allow redistribution of the images on their own or as an asset collection. Projects a user creates are saved in projects/<name>/ and are never committed, because the character art is not Shindo's to redistribute.

Audience

The project reads as a tool for three groups: an illustrator who already has a character and wants a VTuber-style version of it, a streamer who wants a talking face on screen without hiring a rigger, and a small game studio that wants a lightweight dialogue avatar for a visual novel. Node 22.17 or newer runs the editor and a browser renders the result.

The limits worth naming

Mesh Avatar Studio is new, which is why the 119-star figure is dated to the API read on 4 October 2026. The author's README recommends a front-facing, head-and-shoulders PNG with a transparent background as the best input, the hardest of those three constraints to satisfy from stock art. Teams running Claude Code lose the drawn-variant step, which the README states is where blinking and speaking look most natural.

Source

Primary source: shinshin86/mesh-avatar-studio on GitHub.