Video summary

Claude Code Took Over Blender and Created This...

Main summary

Key takeaways

Technology

Technological concepts & product features covered

  • Claude Code + Blender automation via MCP

    • The video claims to connect Claude Code to Blender using an official “Blender MCP” add-on (Python-script driven).
    • The goal is to scale Claude’s Blender abilities by using UI/skill setup and running tasks with minimal friction.
  • Risky “dangerously skip permissions” mode

    • The creator runs Claude with permission-bypass behavior so it won’t ask for approvals during operations.
    • They explicitly note this is not recommended, and that a safer manual-approval mode still works.
  • Toolchain setup workflow (install/config)

    1. Install Blender 5.1.
    2. Install/activate the official Blender MCP add-on in Blender.
      • Verify in Edit → Preferences → Add-ons.
      • Ensure the MCP server is running and auto-start is enabled.
    3. Install/configure Claude Code and connect it to Blender.
    4. Start Claude in a terminal session with “dangerously skip permissions.”
    5. Install Blender MCP server through a provided drag-and-drop MD instruction file inside the Claude Code workspace.
    6. Optionally configure “Full AI” integrations (image generation / 3D generation / vision / segmentation).
      • Requires AI credit replenishment.
    7. Store API keys in a protected .inf file (best practice: keep keys out of plain text).
    8. Restart the Claude session so it detects Blender MCP.
    9. Test connection using a simple command to confirm it “can see my scene.”
  • Core capabilities demonstrated

    • 3D animation generation with physics-like simulation behavior
    • Geometry Nodes editing / parameterized procedural animation
    • World generation pipelines using image/scene-to-3D systems (World Labs / splats)
    • Animation retargeting across different rigs (Akarig ↔ Mixamo)

Key examples / “tests” shown (with outcomes)

1) Character animation: parts falling apart + “tornado” assembly

  • Workflow

    • The character was built (implied via Triple), then split into parts, unioned, and re-textured.
    • Claude runs simulation/animation instructions to create:
      • scattering/fragmenting behavior (parts drop and scatter naturally)
      • a more complex tornado-like animation (parts swirl/assemble)
  • Iteration details

    • The first simulation attempt didn’t behave as desired.
    • Multiple feedback/revision rounds were required for:
      • simulation consistency (proper falling apart)
      • improved tornado assembly behavior
    • Reported effort: 5 iterations total (initial prompt + 4 revisions).
  • Claimed result

    • Produced a “beautiful animation,” and the creator states they couldn’t do that manually as fast.
    • Not perfect, but notably efficient.

2) Geometry Nodes procedural water: animated bubble effect with controls

  • Scenario

    • A water plane in Blender.
    • Claude generates Geometry Nodes to animate the surface so it looks like bubbles.
  • Issues encountered

    • First version produced bubbles too large and too pixelated (scale mismatch vs world units).
  • Debugging loop

    • The creator provided feedback (e.g., “make bubbles ~3x smaller” and improve resolution).
    • Claude adjusted while the creator visually debugged by toggling render/hidden objects to inspect the water effect.
    • Safety claim: it did not delete anything during edits.
  • Final result

    • Bubble behavior controlled by parameters such as:
      • bubble strength
      • frequency
      • spread speed
    • The creator states they didn’t manually edit Geometry Nodes afterward; Claude produced a usable node setup.
    • Also mentions they simplified shading to make results clearer (Cycles render context).

3) World generation with an “image blast”-like pipeline (World Labs, splats/ghosts)

  • Reference system

    • The creator saw a tool called “image blast” (Claude Code skill for World Labs + Full AI).
    • They forked it to test generation within Claude Code / Blender tooling.
  • Key technical points

    • Requires Full AI API to generate 3D objects inside a scene.
    • Uses World Labs to create 3D splats / ghosts.
    • Mentions World Labs can now output colliders and real meshes.
  • Constraints

    • World generation can cost credits (creator suggests up to ~$5 per generation).
    • The pipeline renders nicely in web, while rendering in Blender may be more difficult.
  • Additional experiments

    • Created a Japan-style scene; objects appear to “interact” with the environment (at least visually).
    • Tried a similar result in Blender using Claude Vision for control, but reported vision quality issues:
      • harder to revise objects reliably after generation
      • increased cost per world generation
      • sometimes incorrect placements/rotations
  • Switching model/system

    • After recording, they tried the same setup with Claude Code (codecs) (not fully clear, but described as “codecs”).
    • Claimed improvement due to a better vision model / multimodal integration.
  • Claimed economics

    • Experiments cost “almost $8” due to multiple passes (3D model calls, image generation, PBR materials like patina).

4) Animation retargeting: Akarig ↔ Mixamo (FBX) using Claude

  • Goal

    • Rig a character with Akarig (free) and retarget a Mixamo animation (downloaded as FBX).
    • Demonstrated with a Mixamo “twerk dance.”
  • Challenges highlighted

    • Different skeleton structures:
      • Character Creator has more bones than Mixamo
      • different bone rotations and rest poses
    • Prompting pitfall:
      • initial prompt wrongly assumed rest pose similarity
      • correction discovered: rest pose mismatch due to T-pose exported as an animation frame
  • Iteration / feedback

    • Claude performed multiple checks and attempted pose retargeting fixes.
    • Problems included:
      • arm/hand issues
      • one instance where Blender crashed, but state was recovered after restarting
    • Fix approach:
      • when told about crashes, Claude reverted the model to a working state.
  • Final output

    • Claimed ~90% success:
      • body/legs/head retargeted well
      • arms and especially wrists/fingers had issues (mirroring or mismatch)
  • Suggested next step

    • Build an add-on to retarget between two selected skeletons automatically inside Blender.
  • Additional mention

    • References a LinkedIn example of a Blender add-on that tracks a head and creates parallax effect using the same general MCP/Claude setup idea.

Reviews / Guides / Tutorials emphasized

  • Installation guide

    • Step-by-step setup of the Blender MCP add-on + Claude Code connection.
  • Operational guide

    • Recommended verification steps (MCP auto-start)
    • Key storage in .inf
    • Session restart requirements
  • Practical workflows

    • Physics-like destructible/fragmenting animation via prompting + iterative feedback
    • Procedural water/bubbles via Geometry Nodes with controllable parameters
    • Scene/world generation using World Labs splats + Full AI integration
    • Retargeting animations across mismatched rigs with prompt iterations and debugging

Main speakers / sources

  • Main speaker: The video creator/host (speaking in first person throughout; Claude Code is “my” setup)

  • Systems/sources referenced

    • Claude Code / Claude (agent used for Blender/MCP tasks)
    • Official Blender MCP (add-on)
    • World Labs (splat/mesh/ghost world generation)
    • “Image blast” (referenced prior tool/skill; creator forked it)
    • Akarig (rigging tool)
    • Mixamo (animation source; FBX download)
    • Triple (used earlier to create/split parts and textures, per subtitles)
    • Blender (v5.1), Geometry Nodes, Cycles (Blender features referenced)

Rate this summary

Your feedback will help improve summaries.

Improve this summary

Reprocess with a stronger model when the summary feels incomplete or inaccurate.

Pro

Translate summary in another language

Pro

Ask questions to this video

Chat for follow-up questions, clarifications, and source-backed answers.

Coming soon

Share this summary

Original video