Blender MCP with Claude Code: control Blender from an AI agent

What this guide covers

MCP (Model Context Protocol) lets an AI agent like Claude Code call tools on your machine instead of just writing text. The Blender integration on apillms installs an MCP server that exposes Blender's scene as a set of tools: create objects, move the camera, set materials, render, and more. Claude Code connects to that server and can drive Blender directly from a terminal prompt.

This post follows the Blender guide at /integrations?app=blender and explains the pieces: what the installer does, how to write your first prompts, and how to keep the token bill small while you iterate.

MCP basics in one minute

An MCP server is just a small program that speaks a common protocol and offers a list of "tools" (functions) an agent can call. Claude Code (or Claude Desktop, Cursor, Codex) connects to one or more MCP servers and decides, turn by turn, whether to call a tool or just respond in text.

apillms runs its own general MCP server at https://apillms.com/mcp with tools like chat, generate_image, and balance — useful if you want an agent to call other models or check your account from inside a session. The Blender integration is separate: it installs a local MCP server that controls your running Blender instance, and registers it in Claude Code (and other clients) automatically.

The apillms MCP server, for reference

claude mcp add --transport http apillms https://apillms.com/mcp --header "Authorization: Bearer YOUR_KEY"

You don't need this for Blender control itself, but it's handy if the same Claude Code session should also generate images or check your remaining credits while you work on a scene.

Install the Blender MCP server

From the Blender guide, run the one-line installer (macOS shown; the guide also covers Windows):

curl -fsSL 'https://apillms.com/v1/setup-blender?os=mac' | bash

This installs the Blender MCP server and add-on, and registers the server in Claude Desktop, Cursor, Codex, and Claude Code — so whichever client you already use for coding, it's ready without extra config. Your apillms key is filled in on the guide page before you copy the command, so Claude Code can authenticate against https://apillms.com (Anthropic-compatible base URL, no /v1) once the setup script finishes.

After installation, open Blender with the add-on enabled, then start Claude Code in your project folder as usual. The model you use is whatever you've configured for Claude Code — see the Claude Code setup guide if you haven't connected your key yet.

Your first prompts

Once Blender is running and the MCP server is registered, ask Claude Code to do something concrete in the scene. Specific prompts work better than vague ones, because each tool call costs tokens and you don't want three rounds of clarification:

• "Add a cube at the origin, scale it to 2x1x1, and name it Platform."
• "Create a point light above the scene and set up a basic three-point lighting rig."
• "List the objects currently in the scene and their locations."
• "Render the current view at 512x512 and tell me the output path."

Claude Code will call the relevant Blender tool (create object, set light, render, etc.), read back the result, and report what changed. Keep instructions scoped to one or two actions per prompt so you can check the result before asking for the next step — this also keeps each agent turn short.

Cost tips for iterative 3D work

Blender sessions tend to be long and iterative: you ask for a change, look at the render, ask for another tweak. Each round trip is one agent turn — tool calls plus the model's reasoning and response. On apillms, with prompt caching on by default, a typical agent turn (cached context plus a bit of fresh input and output) on $0.274 is priced using cached reads at the lower rate automatically; per the calculator, one such turn on claude-sonnet-5 costs $0.021, so 1,000 turns run $20.99.

If you're doing a long back-and-forth session (scene setup, materials, lighting, re-renders), that per-turn cost adds up across dozens of turns per session. A cheaper model for routine Blender scripting — object placement, naming, simple transforms — is deepseek-v4-flash, where the same agent turn costs $0.006 (1,000 turns = $5.88). Switch to a stronger model like claude-opus-5-5 only for harder spatial reasoning or when a scene needs careful multi-step planning.

Because prompt caching is on automatically with per-user cache keys, repeated context in a long Blender session — your system prompt, prior scene state the agent re-reads — is billed at the cache-read price rather than full input price. That's part of why the per-turn figures above stay low even as context grows. See prompt caching explained for more on how that works. For a general breakdown of where Claude Code's tokens go, read why Claude Code burns tokens.

Checking the numbers against OpenRouter

If you're used to routing through OpenRouter, apillms lists the same model ids with its own pricing. Compare current rates on /compare or see raw per-pack token estimates on /pricing before committing to a pack.

Takeaway

Install the Blender MCP server with the one-line command, connect it to Claude Code (already registered for you), and start with small, specific prompts. Use a cheap model for routine object and scene edits, and reserve a stronger model for genuinely hard 3D reasoning — the calculator numbers above make the trade-off concrete rather than guessed.

FAQ

Do I need a separate apillms API key for Blender?

No. The Blender installer fills in your existing apillms key during setup, and Claude Code uses the same Anthropic-compatible base URL (https://apillms.com) you already use for coding.

Which Claude Code clients get the Blender MCP server registered?

The installer registers the Blender MCP server in Claude Desktop, Cursor, Codex, and Claude Code, so it's available in whichever one you open.

Which model should I use for routine Blender scripting?

deepseek-v4-flash is the cheapest option in the calculator table ($0.006 per agent turn) and works well for simple object placement and transforms; switch to a stronger model for complex spatial reasoning.

Does prompt caching help in long Blender sessions?

Yes. Caching is on automatically, so repeated context like your system prompt and prior scene state is billed at the lower cache-read price instead of full input price on every turn.