Animato

When to use this skill

• Animate a rigged .fbx / .gltf character from a text request without opening Blender.
• Run the animation loop unattended from a key (ANIMATO_API_KEY / GEMINI_API_KEY / OPENAI_API_KEY) instead of copy-pasting prompts into a chat window.
• Review or gate an LLM-written bpy script before a server executes it.
• Operate, debug, or health-check a local Animato server and its /api/* endpoints.

Route out when the job is rig authoring or mocap retargeting (use Blender directly), real-time gameplay effects (game-vfx), or browser playback of the finished clip (threejs-animation, threejs-loaders).

Instructions

Step 1: Capture one animation packet

Record before any call: the model file and its format, whether it is already uploaded, the motion request in one sentence, the provider/model/key to spend, and whether the source file is expendable. The generated script overwrites the uploaded model in place — if the source is irreplaceable, copy it before uploading. .obj has no skeleton and cannot be animated.

Step 2: Confirm the server and the key

uv run fastapi run main.py   # in the Animato checkout, serves UI + API on :8000
python3 .agent-skills/animato/scripts/animato_agent.py doctor

doctor checks server reachability, key presence (never printing the value), and the gate script. Configure with ANIMATO_SERVER, ANIMATO_PROVIDER (gemini | openai), ANIMATO_LLM_ENDPOINT, and ANIMATO_MODEL; see references/agent-loop.md.

Step 3: Run the gated loop, not the raw endpoint

export ANIMATO_API_KEY=...   # free-tier key is enough; one inference per animation
python3 .agent-skills/animato/scripts/animato_agent.py animate \
  --file "./X Bot.fbx" --message "wave hello with the right arm, 2 seconds"

--mode local (default) does upload → /api/prompt → one LLM call → static gate → /api/run, writing the prompt and the generated script into --out-dir (default animato-out/) so a failure is a reviewable artifact. Use --dry-run to stop before execution, and --mode server only when you accept that /api/chat executes the script before any gate can run.

Step 4: Read the gate output before re-prompting

A rejected script means the request or the model choice was wrong, not that the loop should retry.

python3 .agent-skills/animato/scripts/validate_bpy_script.py animato-out/X-Bot.generated.py \
  --model-path public/upload/X-Bot.fbx

The gate refuses stale Blender APIs, missing keyframes, a missing frame range, a missing export, export_animations=True / bake_anim=True omissions, format mismatches, and host-side calls such as subprocess/os.system/eval. Full table: references/bpy-script-contract.md.

Step 5: Verify the animated file, not just the HTTP status

/api/run returns HTTP 200 even for a failed script — branch on ok, then confirm output_url loads and actually carries a clip (gltf.animations / FBX .animations non-empty, mixer.update(delta) called every frame). Fix a wrong clip with a fresh request; drop a bad clip deterministically:

python3 .agent-skills/animato/scripts/animato_agent.py remove --filename X-Bot.fbx --name wave

Step 6: Keep the loop honest

Treat /api/run and /api/chat as a local remote-code-execution surface: trusted machine only, never public without a sandbox. Never spend more than one inference per animation on autopilot, and never claim an animation works from a passing gate alone — the gate is static.

Examples

Verify the whole loop offline (no key, no Blender)

python3 .agent-skills/animato/scripts/selftest.py

Runs a stub Animato server and stub LLM through six cases: happy path, gate rejection blocking execution, --dry-run, /api/chat mode, missing-key failure, and doctor.

Animate a model that is already uploaded, with an OpenAI-compatible provider

ANIMATO_PROVIDER=openai ANIMATO_LLM_ENDPOINT=https://api.example.com/v1 ANIMATO_MODEL=my-model \
python3 .agent-skills/animato/scripts/animato_agent.py animate \
  --filename X-Bot.fbx --message "idle breathing loop, 3 seconds" --dry-run

Inspect the prompt the server builds

python3 .agent-skills/animato/scripts/animato_agent.py prompt \
  --filename X-Bot.fbx --message "wave hello" --out prompt.txt

Best practices

1. Copy the source model before uploading — the animated export replaces it in place.
2. Prefer --mode local: gate the script before execution instead of after.
3. Spend one inference per animation; fix the request or the prompt, do not loop the model.
4. Take bone names, axes, and units from the /api/prompt dump, never from an example script.
5. Keep the persisted prompt and generated script together so a bad result is diagnosable.
6. Keep the server local and trusted; keys stay in the environment, out of logs and artifacts.
7. Confirm the exported file plays before reporting success.

References

• references/agent-loop.md — endpoints, modes, environment variables, failure routing.
• references/bpy-script-contract.md — gate rules and removed-API table.
• references/example-bpy-script.py — gate-passing script shape.
• references/upstream.md — provenance, requirements, safety boundary.
• otdnnc/Animato