How to use the Blender Scene CLI agent skill

What does the Blender Scene CLI skill make the agent do?

An agent skill is a folder with a SKILL.md file that tells an AI agent how to perform a task with a specific tool. This skill covers Blender, a free 3D modelling and rendering program, through the cli-anything-blender command-line interface.

With the skill, the agent can add 3D primitive objects, create and assign materials, add cameras and lights, set animation keyframes, save the scene as a JSON project file, undo or redo changes, and render images or previews. The tool generates Blender Python (bpy) scripts and runs Blender to produce the actual render.

• Objects: add, duplicate, transform, rename, set parents and remove
• Materials: create, assign and set properties such as colour, metallic and roughness
• Cameras and lights: add, set properties and choose the active camera
• Animation: set and remove keyframes, frame range and FPS
• Rendering and previews: render the scene or capture preview images

Install the Blender Scene CLI skill

You can install the skill in one of two ways. Download the ZIP from the skill page on this site and unzip it into ~/.claude/skills/ for your user account, or into a project's .claude/skills/ folder for one project only.

If you use the Claude apps, you can instead upload the ZIP under Settings → Capabilities → Skills.

The skill also relies on the command-line tool itself, which is installed separately with pip. Blender must be installed on your system before the tool can render.

• Python 3.10 or newer
• Blender 4.2 or newer installed on the system
• pip package cli-anything-blender

How do you trigger the skill with natural requests?

You do not need to type commands yourself. Describe the change you want in plain language and mention Blender, and the agent will choose the matching cli-anything-blender commands. Good requests name the object, the change and, if relevant, the output you want.

If you want to check your own setup, you can run the tool directly. Running it with --help lists the commands, and running it with no subcommand opens an interactive session with tab-completion and history.

• Add a red cube at the origin and give it a metallic material.
• Put a camera looking at the cube and a sun light above it, then render a quick preview.
• Move the cube up 2 units and add a keyframe at frame 1 and frame 48.
• Undo the last change and show me the scene info.

Examples of what you can ask for

To start a project, the agent creates a scene file with a .blend-cli.json name. Agents are told to pass --json for machine-readable output, so you can see the scene state the agent reads.

For visual checks, the agent can capture a preview bundle. The source notes that the quick recipe produces hero.png from Eevee preview rendering and workbench.png from Blender Workbench rendering.

What about safety, scripts and network use?

The skill runs scripts: the tool generates Blender Python scripts and runs Blender with them to render. Review what the agent is about to run if you are working on important files. The skill does not use the network, according to its listing.

The source requires absolute paths for all file operations, including rendering and project files, because relative paths can fail in background execution. Keep your project files in a folder you control and back them up before large automated edits.

• Scripts are executed, so run the skill on scenes and folders you are happy to let the tool write to.
• Undo and redo cover up to 50 levels of history, but save copies of important .blend-cli.json files anyway.
• Use absolute paths for projects and outputs.

Tips for better results

Keep scenes simple. The skill covers a defined command set rather than Blender's full GUI tools, so complex modelling is outside its scope.

Ask the agent to capture a preview after changes instead of judging the scene from JSON alone. The source says agents should read the returned image file paths rather than expect image data inline.

Use the session commands to step back through changes while you experiment.

• Name objects clearly so the agent can refer to them by name.
• Ask for a preview after each major change.
• Use session status and history to confirm what has happened so far.