Skip to content
Stable3D & Animation

MCP for Blender — Connect AI Agents to Blender

Give your agent a real Blender session instead of a blank Python prompt. Every tool call is typed, validated against the current scene, and written to an audit log.

Read-only by default8 typed toolsCheckpointed writes
Technical specification
Host software
Blender 4.2 LTS and newer (4.5, 5.x)
Runtime
Python 3.11+, bundled with Blender
Transport
stdio (local) and streamable HTTP (remote studio)
Install
Blender addon + npx dccmcp-blender
Modes
Interactive viewport, background/headless, batch farm workers
Platforms
macOS, Windows, Linux
Last reviewed
2026-09-20
Overview

What is MCP for Blender?

MCP for Blender is a local MCP server plus a Blender plugin that connects AI agents to a real Blender session. Agents discover 8 typed tools — including scene.inspect, object.create, mesh.edit and geometry_nodes.set_input — and call them to inspect, create, modify and validate work inside the live application.

Unlike generated Blender scripts, every call is validated against your current document, classified as read, write or execute, and recorded in an audit log. Blender stays open, the document stays live, and destructive operations are checkpointed so they can be rolled back.

It works with Claude Code, Codex, Cursor, GitHub Copilot, Claude Desktop, Gemini CLI, Trae, Doubao, Qwen Code, CodeBuddy, Custom MCP clients, Windsurf over stdio or streamable HTTP.

  • Host software: Blender 4.2 LTS and newer (4.5, 5.x)
  • Runtime: Python 3.11+, bundled with Blender
  • Transport: stdio (local) and streamable HTTP (remote studio)
Why it breaks

What goes wrong when agents drive Blender without a contract

01

Generated scripts assume a scene that does not exist

An LLM writing bpy code has never seen your file. It guesses collection names, object paths and modifier stacks, then fails halfway through and leaves a half-modified scene behind.

02

A crash wipes out an hour of work

Scripts that call destructive operators without a checkpoint can corrupt a scene, break linked libraries, or take the viewport down with them.

03

Nobody can review what the model did

When an automation runs unattended, there is no record of which object was changed, by whom, at what time — which makes studio sign-off impossible.

How it works

How MCP for Blender keeps intent, structure and traceability intact

A local MCP server plus Blender addon that exposes typed, audited tools for scene inspection, modeling, geometry nodes, shading, rendering and headless batch jobs.

Inspect before it acts

The agent calls scene.inspect and receives a structured summary of collections, objects, modifiers, materials and units before it proposes any change.

Typed tools instead of free-form bpy

Tools accept validated schemas — mesh names, unit-aware dimensions, slot indices. Invalid calls are rejected at the edge, not in the middle of a render.

Checkpointed write operations

Every mutating tool runs inside a transaction that can be previewed, applied, or rolled back. Undo history stays intact for manual cleanup.

Batch and headless parity

The same tool surface drives Blender in background mode, so an agent can build variants across a render farm with identical semantics.

Tool reference

MCP for Blender tools — every tool and its risk class

Agents discover this list at connect time. You decide which entries are auto-approved, which need a human and which are denied in production files.
  • scene.inspectStructured snapshot of collections, objects, transforms, modifiers, materials and units.read
  • object.createCreate primitive or data-driven objects with validated transforms and parenting.write
  • mesh.editNon-destructive mesh operations with explicit selection scoping and dry-run preview.write
  • geometry_nodes.set_inputDrive geometry node inputs by socket name and type, including simulation zones.write
  • material.assignAssign or build shader graphs, including principled BSDF parameter sets.write
  • render.captureQueue viewport or final renders and return rendered images for verification.execute
  • pipeline.checkpointCreate, list and restore named checkpoints of the current file state.write
  • asset.searchSearch local asset libraries and return typed asset references.read
Install

Install MCP for Blender in four steps

Everything runs locally. The server binds to loopback by default and only exposes the tools you approve in Blender.

Pre-release. These commands describe the interface we are shipping for MCP for Blender — the packages are not published yet. Get notified when MCP for Blender launches.

STEP 01

Install the addon

Add the DCCMCP addon to Blender and enable it in Preferences → Add-ons. The addon opens a local control channel and never listens on a public interface by default.

blender --command extension install dccmcp_blender
STEP 02

Start the MCP server

Run the server from your project directory. It discovers a running Blender session and advertises its tool list to the client.

npx dccmcp-blender@latest serve --port 7331
STEP 03

Register with your agent

Register the server with Claude Code, Codex, Cursor, OpenClaw, VS Code or any other MCP-compatible client — they all speak the same stdio or HTTP transport.

{
  "mcpServers": {
    "blender": { "command": "npx", "args": ["dccmcp-blender", "serve"] }
  }
}
STEP 04

Set a policy

Choose which tools are auto-approved, which require confirmation, and which are disabled in production files. Defaults are read-only.

dccmcp policy set --profile studio-default

MCP for Blender works with the agents you already use

  • Claude Code
  • Codex
  • Cursor
  • GitHub Copilot
  • Claude Desktop
  • Gemini CLI
  • Trae
  • Doubao
  • Qwen Code
  • CodeBuddy
  • Custom MCP clients
  • Windsurf

Any other MCP-compatible client works too — one config entry, stdio by default.

Read the docs
FAQ

Questions teams ask before they install

If something is missing, ask us directly — we answer technical questions with technical answers.
  • No. The server runs on your machine and talks to your local Blender process. Scene data only leaves your network if you explicitly configure a remote HTTP transport for a studio deployment — and even then, only the tool results the agent asked for are transmitted.

  • Mutating tools run behind a policy layer, and the default profile is read-only. Studio and Enterprise add checkpoints before every destructive operation, so a bad call can be rolled back through the host application's undo stack plus DCCMCP. On the free Community edition the protection is the read-only default and the approval gate — checkpointing and rollback are paid features.

  • Yes. The same tool surface is available in -b/--background sessions, which is how most studios drive variant generation and render farms.

  • Blender 4.2 LTS and newer, including 4.5 LTS and the 5.x line. Legacy 3.x support is available on the enterprise track.

  • Generated scripts are unvalidated, unrunnable in isolation and impossible to review. DCCMCP exposes a stable, discoverable tool surface with schemas, permissioning and audit trails, so the agent works against a contract instead of a guess.

npx dccmcp-blender@latest serve

Put an agent inside your real tools — without putting your pipeline at risk

Install an integration, set a policy, and let your agent inspect, build and validate work in Blender, Maya, Houdini, 3ds Max, Rhino, ZBrush, Photoshop, FreeCAD, QGIS or OpenCV. Free to start, local by default.

macOS · Windows · Linux — no cloud dependency required