Claude Code Blender skills: 24 read, one tested in headless Blender, and plain Haiku already got the models right

Summary

If you want Claude Code to build things in Blender, pick a skill by how it reaches Blender first. Half of the 24 we read run headless, a quarter need a live Blender MCP connection, and five only describe a look. For simple props you may not need one at all.

The reason is simple: we ran it. We sorted every Blender skill in our skills directory, then gave Haiku six modeling tasks with exact sizes, from a table to a slab with a hole, with and without EL4CTEO’s blender-core. We ran each script in headless Blender 5.1 and measured the exported model:

• Plain Haiku got all 18 right. Every script ran, exported a GLB, matched the requested size within 2%, and came out closed with no open edges.
• With blender-core, 16 of 18. Two scripts crashed on wrong argument names: use_selected for the glTF exporter, which has always been use_selection, and an old diameter argument for an icosphere.
• The skill changed how the scripts work. 94% printed their own measurements, such as dimensions and triangle counts, against 6% without it. All but one built geometry through bmesh, as its rules ask, and none used bpy.ops primitives.
• It cost more. Scripts were about 45% longer in tokens, and the 18 runs cost 40% more.

Our picks

1. Choose by connection. No Blender MCP server? Use one of the 12 that run headless. If you already run one, the MCP skills can look at the scene and screenshot it.
2. Skip the skill for simple props. For boxes, stairs, shelves and booleans, plain Haiku wrote working Blender 5.1 scripts every time. Write exact sizes in meters and say Z is up.
3. Use blender-core when the agent must check its own work. Its rule is numbers, not adjectives: report triangles, objects and dimensions before saying done. That is what you want in a long, multi-step build where a wrong step can hide.
4. Pair an art direction skill with one that runs code. The five style skills from arjun988, such as anime and found-footage horror, are short notes on look and lighting. They run nothing on their own.
5. Check the version and the bill. EL4CTEO’s set targets Blender 5.2 or newer; OpenMontage’s 3D skill names 4.5. That one calls paid services: Tripo through Atlas Cloud at $0.22 to $0.44 a model, Hunyuan 3D through fal at $0.225.

What will you hand to the agent: one prop, or a scene it has to build and check over many steps? The second is where a skill pays for itself.

How we tested

1. The catalog. Every skill in AgentCraft’s catalog whose name or description mentions Blender: 24 skills in 10 repos. We read each SKILL.md on November 7, 2026 and noted how it runs Blender, the version it names and anything it needs, such as an API key.
2. The test. Six tasks with exact sizes: a table of 5 objects, a 10-step staircase, a bookshelf with an Oak material, a mug with a handle, a two-material tree, and a 2 m slab with a 0.4 m hole cut through it.
3. Two arms. No skill, and blender-core in the project with the prompt starting /blender-core. Each task three times per arm, each a fresh claude -p session on Claude Code 2.1.314 with --model haiku, no tools: the model only wrote the script.
4. Running it. We ran every script in Blender 5.1.2 with -b --factory-startup and a 120 second limit, then imported the exported GLB into a fresh Blender and measured overall size, mesh objects, open edges (after welding the seams glTF splits), materials and, for the slab, volume. We checked the measurer on a correct slab (volume 0.18744 against 0.18743 expected, 0 open edges) and on an open plane (4 open edges).
5. One fix. On the first pass every tree failed on width. An icosphere of radius 0.6 is only about 1.14 m wide along X, so the check was wrong, not the trees; we dropped that axis for both arms.

Results

How the 24 skills reach Blender

Route	Skills	Examples
Needs a Blender MCP connection	6	arjun988’s modeler, director, archviz and asset optimization; RobLe3’s wireframe-to-3d; Scenario’s Blender extension, which also needs a Scenario API key
Headless or MCP	7	EL4CTEO’s core, asset QA, animation and rigging; jangtrinh’s agent core; Scenario’s 2D animation; jev-blender
Headless scripts only	5	EL4CTEO’s add-on development, a glTF web pipeline, DESIGN:OS parametric parts, a holographic card builder, a game sprite pipeline
Text guidance only	5	arjun988’s anime, found-footage horror, apocalypse, biome and body horror styles
Calls paid services	1	OpenMontage’s 3D asset generation

The test, 18 runs per arm

	No skill	blender-core
Script ran and exported a GLB	18	16
Size within 2%, task checks passed	18	16
No open edges	18	16
Prints its own measurements	1	17
Builds through bmesh	8	17
Uses bpy.ops mesh primitives	6	0
Median script	79 lines, 3,805 output tokens	85 lines, 5,508 output tokens
Cost, 18 runs	$0.070	$0.098

Both failures stopped before export. The bookshelf script built all eight boards and printed their sizes, then passed use_selected to the glTF exporter. The tree script called bmesh.ops.create_icosphere with diameter, a name Blender dropped years ago for radius. Building through bmesh, as the skill asks, means more arguments to remember, and Haiku remembered two wrong.

Which repos are growing? From our catalog’s star counts on October 29 and November 6, arjun988’s set gained 451 stars, 29% in 8 days, the fastest of the dedicated Blender repos. EL4CTEO’s gained 160. Scenario’s skills repo gained 2,254, but most of it is not about Blender.

What to ignore

• Skill count per repo. arjun988 lists nine Blender skills, but five of them are a page or less of style notes.
• Version labels as a hard stop. blender-core says 5.2 or newer; its rules worked in 5.1, and neither failure came from a version change.
• Our tasks as a ceiling. These are props a model can build in one script. Rigging, animation and look development, where the bigger skills spend most of their words, were not tested.

Limitations

• Haiku only, one script per task. A larger model, or an agent allowed to run Blender and fix its own errors, would likely close the two failures.
• Only blender-core was tested. The MCP skills need a live Blender session, which a batch test cannot fairly drive.
• The catalog is ours. It lists repos with 100 or more stars whose skills passed our review. Blender skills inside larger projects that we did not list are missing.
• Reading, not running, for 23 of 24. Routes and needs come from each SKILL.md as written on November 7, 2026.

Reproduce it

The script is claude-blender-skills-2026-11.py. With Claude Code logged in and Blender installed, run python3 claude-blender-skills-2026-11.py run <folder> <blender-core SKILL.md> <path to Blender>, then score <folder>. The data, claude-blender-skills-2026-11.json, has every catalog row with its route, needs, version and star counts, and every test run with its measurements, errors and cost.

Sources

1. EL4CTEO/blender-skills: blender-core · official
2. arjun988/blender-skills · official
3. Blender Python API · official
4. AgentCraft: Claude Code skills directory · blog