Text-to-Anim

Free · Open source · 100% local

Type a sentence.
Get a Unity animation.

Generate mocap-style Humanoid clips from text, on your own GPU. No cloud, no credits, no subscription. Drop the result on any Mixamo character, or on a whole crowd.

▶ Watch the demo How to set it up
  • Unity 6 · Humanoid
  • AMD · NVIDIA · Intel (Vulkan)
  • Works with Claude Code via MCP
› a zombie shambles forward slowly, dragging its feet, arms reaching forward
🆓

Free & open source

The plugin and MCP server are open source. Generation runs open-weight models (NVIDIA Kimodo via kimodo.cpp) locally. Make a thousand clips, and it costs nothing.

🧍‍♀️🧍🧍‍♂️

Any humanoid, any crowd

Clips come out as standard Unity Humanoid animations, so they retarget onto Mixamo, VRM, Synty or your own rigs. Change the seed for natural variation across a crowd.

🎬

Directed by your AI agent

Connect Claude Code (or any MCP client). Describe a scene and it plans the shots, generates each one, checks the previews and redoes the bad takes.

See it in action

Every animation below was generated from a one-line prompt on an 8 GB AMD RX 6600 XT. Nothing was hand-keyed or motion captured.

Crowds. A few prompts, many seeds. No two people move the same.
NPC moments. "a person waves hello with their right hand"
The fun ones too. "a person moonwalks backward, sliding their feet smoothly along the floor…"
Any Mixamo character. One clip ("a person throws punches like a boxer") on six different rigs.

How it works

  1. 1

    You describe it

    "Guard paces between the pillars, hears a noise, turns."

  2. 2

    Your agent directs

    Claude Code splits it into shots and calls the local MCP tool for each one.

  3. 3

    Your GPU animates

    A motion-diffusion model generates every take locally, in about 20 s per 5 s clip on a mid-range card.

  4. 4

    Unity gets a clip

    A Humanoid AnimationClip lands in your project, auto-checked for facing, feet and retargeting.

Why an agent? Motion models are great at moving, but you still have to plan and judge the takes. Claude reads the scene, writes motion-capture-style prompts, looks at each preview and regenerates the bad ones. You stay the director without babysitting every take.

Setup

Early preview: Windows, Unity 6, and a Vulkan GPU with 8 GB VRAM.

1

Install

git clone <repo>  &&  .\setup.ps1

Downloads the engine and models (~6 GB) after you accept the model licenses.

2

Add to Unity

Package Manager → Add from disk →
com.local.text-to-anim

Any .kimodo file imports as a Humanoid clip.

3

Connect your agent

claude mcp add kimodo-motion -- \
  .venv\Scripts\python mcp_server\server.py

Then: "make a crowd of 20 people waiting for a bus".

The public repo and one-command installer are on the way. Follow along to get notified.

FAQ

Is it really free?

Yes. Everything runs on your machine, so there are no credits or subscription. The plugin is open source. The models are free open-weight downloads under their own licenses: the NVIDIA Open Model License for the motion model, and the Meta Llama 3 Community License for the text encoder. Read them before shipping commercially.

What hardware do I need?

A Vulkan-capable GPU with about 8 GB VRAM (AMD, NVIDIA or Intel) and 16 GB RAM. On an RX 6600 XT, a 5 s clip takes about 20 s, and faster cards are quicker.

Does it work with Mixamo characters?

Yes. Clips are standard Unity Humanoid animations, so any character set up as Humanoid works. Every generated clip is automatically checked on more than one rig.

Do I need Claude?

No. You can generate clips directly. An MCP-capable agent like Claude Code is just the nicest way to direct multi-shot scenes. It uses your own Claude plan, and the plugin never touches your credentials.

What can't it do (yet)?

No finger or face animation, no exact targeting yet (e.g. "sit on this chair"), and prompt-following isn't perfect, so you'll sometimes regenerate. Clips max out at 10 s per segment, but segments chain together.