Claude Opus 5.5 + Blender MCP · 3D Modeling, Rigging and Animation

Quick Summary

• Aidan Stanik let Claude Opus 5.5 operate an open Blender scene through MCP, writing the Blender Python used to create, rig and animate the examples shown in the video.
• The creator reports a 1.4-million-polygon space station and a creature with 115 bones. These are creator-reported project measurements, not independent benchmarks of every Claude and Blender workflow.
• No generative video model produced the frames. The scene, camera, materials, rig and animation remained inside Blender, which makes individual objects, bones, keyframes and render settings editable after generation.
• The strongest result is a creature that breathes, blinks, moves its fingers and tracks a flying bug. The video also exposes defects: fingers intersect a branch, a high-resolution render develops a black material glitch, and the caterpillar feels too toy-like to the creator.
• Blender now offers an official connector built by Blender's developers. A popular community project, MCP for Blender, is a separate third-party integration. Pick one setup path and verify exactly which server you are installing.
• The reliable production pattern is staged: secure the project, block the scene, approve stills, audit geometry, build the rig, test deformations, render a short preview, then commit to the expensive final output.

What Claude built during the experiment

The opening montage shows three types of output: a complex space environment, a rigged alien-lizard scene and a soft, bouncing caterpillar. Aidan says he supplied references and natural-language direction while Claude wrote the code, built the scenes and rendered the results from Blender.

The headline numbers are striking, but should be read in context. Aidan reports that the space station contains 1.4 million polygons and the creature rig contains 115 bones. Polygon and bone counts describe complexity, not quality. A lighter model with clean topology and a purposeful control rig may be more useful in production than a denser model.

The experiment is best understood as a creator test, not a controlled benchmark. It demonstrates that one skilled user obtained these outputs in a day with a particular model, connector, prompt history, machine and set of references. It does not establish the same result for every scene or user.

What no AI video model actually means

Claude is still an AI model, but it is not synthesizing every final frame through a video diffusion model. MCP gives the model callable tools, and the Blender connector exposes the open scene through Blender's Python API. Claude can inspect the scene, write code, change objects and request views or renders.

The final pixels are produced by Blender's viewport or render engine from persistent geometry, materials, lights, cameras, rigs and keyframes. This matters because the underlying scene remains addressable. You can move one finger, change one light, retime one keyframe or replace one material without asking a video model to reinterpret the whole shot.

Deterministic does not mean effortless. The Blender file can still contain bad topology, inefficient modifiers, broken weight painting, intersecting meshes, unstable simulations or expensive shaders. Agent control changes who performs the operations, not the physics and engineering of 3D production.

Why Aidan temporarily skips generative video

Aidan contrasts this test with his earlier workflow of building Blender previs and sending it to a generative video model. His stated reasons are cost, repeated prompting and visual inconsistency between outputs. Keeping the shot inside Blender preserves the same models, camera, spatial relationships and animation controls across iterations.

This does not make Blender a universal replacement for generative video. Blender gives precise control but requires 3D assets, lighting, rigging, rendering and technical review. Video models remain useful when the desired result is organic, photorealistic or faster to explore than a complete 3D build.

A practical hybrid is still valid: use Blender for layout, continuity, camera and hero objects, then decide whether the final look should remain 3D, receive compositing and grading, or become controlled input for a separate video model. The video deliberately stops before that final branch to test Blender on its own.

First choose the correct Blender MCP path

The phrase Blender MCP now refers to more than one implementation. Anthropic's current documentation says Blender's developers built an official connector that works through Blender's Python API. The popular GitHub repository MCP for Blender, formerly blender-mcp, explicitly describes itself as a community plugin not affiliated with the Blender Foundation.

The transcript shows a connected Blender entry in Claude Code's /mcp panel, but the narration does not name the repository or package. For that reason, this guide does not claim that Aidan used a specific implementation. Follow the setup that matches the connector or repository you intentionally selected.

Do not install several Blender servers under similar names and assume they are interchangeable. Duplicate local servers can compete for the same Blender session, expose different tools and make troubleshooting much harder.

Current official connector setup

Claude Academy lists Claude Desktop and Blender 4.2 or later as prerequisites. In Claude Desktop, open Customize, choose Connectors, search for Blender and add it. Then open Blender's MCP Server page, install the Blender add-on as instructed, open Blender Preferences, enable the BlenderMCP add-on and start the MCP server from inside Blender.

Anthropic says the connector requires Blender and Claude to run on the same machine. It operates on the open scene through the Python API, and the changes are written to disk when you save the Blender file. Claude Academy also says the connector uses open MCP and can work with other MCP clients, including Claude Code.

In Claude Code, /mcp is the status panel used in Aidan's video. The official Claude Code documentation says it lists configured servers and their connection health. Confirm Blender is connected before issuing a build request, and inspect the tool names so you know which implementation is active.

Community MCP for Blender setup

If you intentionally choose the community project, use its current repository and documentation. The project was renamed from blender-mcp to mcp-for-blender. Existing setups may still use the old package name, while the current quickstart recommends mcp-for-blender.

The repository's current Claude Code command is claude mcp add blender uvx mcp-for-blender. It also provides uvx mcp-for-blender install-addon, after which you enable Interface: MCP for Blender in Blender, open the sidebar with N and confirm that the server is running. Follow the repository rather than copying an old tutorial because package names and installation steps have changed.

This community server can execute Blender Python and import assets from external libraries. Its documentation says arbitrary Python execution is enabled by default and offers BLENDER_MCP_SAFE_MODE=1 to block risky filesystem, process, network and persistent-code behavior while keeping normal Blender operations. Review that boundary before connecting it to client work.

Secure the project before giving an agent control

Start from a duplicate .blend file, not the only copy of a live production. Save numbered checkpoints before asset imports, topology changes, rig generation and batch animation. Keep linked libraries and external textures in a project folder you can restore.

Ask the agent to show a plan before any destructive or expensive operation. Separate read-only inspection from changes. Approve network access, external downloads and file writes deliberately. If the connector offers a safe mode or per-tool approval, use it until the workflow is understood.

Never paste private credentials into scene objects, prompts or scripts. Validate downloaded add-ons and 3D assets, scan their licenses, and keep a provenance record. A free model may still require attribution or forbid commercial use.

The one natural-language prompt

Aidan copies a reference from Blender's website and gives Claude a conversational brief. He asks for a realistic creature scene, says the creature should eventually be rigged and animated smoothly, acknowledges that the build will be difficult, and allows the agent to research a free character model or construct one itself.

The useful part is not that the prompt is unstructured. It communicates the target, future technical requirement, difficulty and fallback path. A strong production version should also define delivery resolution, animation length, render engine, polygon budget, approved asset sources, license requirements and what must be shown before the final rig or render.

Talk naturally, but close the gaps that can become expensive. A junior 3D artist still needs constraints, reference rights, a review schedule and a definition of done.

A safer build order for complex scenes

1. Analyze the reference and write the visual goal. 2. Block the scene with primitive geometry and one camera. 3. Approve composition and scale. 4. Build or import the hero asset. 5. Audit topology, modifiers and polygon count. 6. Define the rig and required actions. 7. Test extreme poses and deformations. 8. Animate a short representative shot. 9. Add materials, lighting and depth of field. 10. Render low resolution, fix defects, then render the final output.

Ask Claude to stop at each gate. The agent can provide a contact sheet, scene summary, object counts and a list of unresolved issues. This prevents it from spending hours polishing a camera angle or rig that was never approved.

For a creature, define the actions before the skeleton: breathing, blinking, finger tapping, throat expansion, head tracking, antenna follow-through and idle weight shifts. The rig only needs controls that support the shot and future reuse.

Rigging review: bones are only the beginning

A rig must be judged by deformation and control, not by bone count. Inspect joint placement, parent hierarchy, constraints, inverse kinematics, weight distribution, facial controls, scale application and whether named controls are understandable to a human artist.

Run a pose test before animation. Rotate every major joint to its expected limit, curl the fingers, bend the tail, open and close the mouth or throat, and move the antennae. Check for collapsing volume, candy-wrapper twisting, disconnected surfaces and geometry penetrating the environment.

The video shows why this matters. Claude's creature animation contains a finger intersecting a branch. Aidan corrects the pose manually because the rig remains editable. That is a workflow success, but it also proves that the agent did not deliver a fully approved final shot without human intervention.

Animation review: inspect motion, not only hero frames

The creature's small actions sell the result: breathing, blinking, finger movement, throat motion, head tracking and antenna follow-through. Secondary motion and reaction make the shot feel observed rather than mechanically keyframed.

Review the animation as a frame strip and at normal playback speed. Look for foot or hand sliding, sudden speed changes, gimbal flips, constraint pops, camera clipping, interpenetration, disappearing objects and motion that becomes unreadable behind depth of field.

Keep the initial proof short. Aidan says the 480p version took roughly one to two hours to render, while a later 1080 output ran for about two hours and revealed a severe material or shading glitch halfway through. A small preview is not only faster. It is a risk-control tool.

Rendering review: separate scene quality from render quality

A strong viewport composition can still fail during a final render. Verify frame range, resolution, color management, denoising, motion blur, samples, transparent materials, textures, linked data and whether the selected engine supports every shader feature used in the scene.

The caterpillar example is shown in Eevee rather than Cycles. That tradeoff favors speed, but every project should choose its render engine intentionally. Use the engine that meets the look, hardware and deadline rather than assuming the slower option is automatically better.

Render three diagnostic outputs before the final movie: a low-sample beauty pass, a material or object-ID check, and representative frames from the beginning, middle and end. The middle-frame check would likely have caught the creature turning black before the two-hour render completed.

A practical human approval scorecard

Scene: correct scale, no duplicate objects, sensible naming and acceptable polygon budget. Model: clean silhouette, usable topology, no hidden intersections. Rig: clear controls, stable constraints, tested deformations. Animation: readable timing, grounded motion, no clipping or pops. Look: consistent materials, lighting and depth of field. Render: stable across the full frame range, correct resolution and acceptable noise.

Mark every category pass, fix or reject. A fix is local and preserves the shot. A reject means the model, reference, rig strategy or animation concept needs to be replaced. This prevents a technically complicated scene from passing merely because it looks impressive in one still.

Archive the prompt, scripts, asset sources, Blender version, connector version, render settings and final .blend file. Agent-generated work is easiest to maintain when the production record is as reproducible as the scene.

Where this workflow fits best

Claude plus Blender MCP is strongest for procedural environments, product visualization, stylized characters, motion graphics, previs, camera blocking, asset cleanup and repetitive scene operations. These tasks reward precise scene access and code-driven repetition.

It is weaker when the brief demands production-ready organic character topology, nuanced acting, complex cloth or hair simulation, studio-grade facial animation, or guaranteed final quality without specialist review. The agent may accelerate the first 80 percent while the last 20 percent still needs an experienced modeler, rigger, animator, lighting artist or technical director.

The real change is not that Blender runs itself. It is that natural language can now drive a large part of Blender's existing technical surface. The value comes from combining that speed with deliberate approvals, backups and human taste.