Blender MCP Server for n8n

A Model Context Protocol (MCP) server that exposes Blender's 3D modeling capabilities to n8n workflows.

System Architecture

To avoid confusion, this project consists of two core components:

Blender MCP Addon: A plugin installed inside Blender. It acts as the local execution engine, receiving commands and manipulating the 3D scene.
MCP Bridge Server: A standalone Python server (src/) that acts as the gateway. Clients like n8n connect to this Bridge, which then forwards commands to the active Blender Addon.

graph LR
    n8n[n8n / AI Agent] -- "MCP (HTTP Streamable)" --> Bridge[MCP Bridge Server]
    Bridge -- "Local WebSockets" --> Addon[Blender MCP Addon]
    Addon -- "Python API" --> Blender[Blender Engine]

Quick Start

1. Install Dependencies

We recommend using uv for fast virtual environment management and package installation:

# 1. Sync dependencies (automatically creates .venv if missing)
uv sync

# 2. Activate it
# Windows:
.venv\Scripts\activate
# macOS/Linux:
source .venv/bin/activate

Configuration

Create a .env file in the root directory to customize your setup:

Variable	Description	Default
`MCP_BRIDGE_HOST`	Host IP for the MCP Bridge Server	`0.0.0.0`
`MCP_BRIDGE_PORT`	Port for the MCP Bridge Server (n8n connects here)	`8008`
`BLENDER_ADDON_HOST`	IP where Blender is running (for Bridge to connect to)	`127.0.0.1`
`BLENDER_ADDON_PORT`	Port Blender addon is listening on	`8888`
`BLENDER_ASSETS_DIR`	Directory to resolve relative textures/HDRIs	(Optional)

Installation

Method 1: Zip & Install (Recommended)

Zip the blender_mcp_addon folder (into blender_mcp_addon.zip).
Open Blender.
For Blender 4.2+: Go to Edit > Preferences > Get Extensions > Click the dropdown arrow in the top right > Install from Disk... and select the .zip. For Blender 4.0 / 4.1: Go to Edit > Preferences > Add-ons > Install... and select the .zip.
Search for "Blender MCP" and enable the checkbox.

Method 2: Manual Copy (Developer)

Copy the blender_mcp_addon folder to your Blender addons directory:
- Windows: %USERPROFILE%\AppData\Roaming\Blender Foundation\Blender\4.x\scripts\addons
- macOS: ~/Library/Application Support/Blender/4.x/scripts/addons
Restart Blender.
Enable "Blender MCP" in Preferences.

Why a folder instead of a single file?

As the addon grows, a single 1800+ line file becomes unmaintainable. We've split the logic into functional modules (modeling, materials, anim, etc.) to make it professional, readable, and easier to extend.

Usage

1. Start Blender MCP Addon

Open the N Panel (press N in the 3D Viewport).
Look for the Blender MCP tab.
Click Start MCP Server.

2. Start the MCP Bridge Server

# Standard mode
uv run python -m src.main serve

# Recording mode (Save all commands to a file)
uv run python -m src.main serve --record my_session.json --name "Building My House"

The server will start on http://localhost:8008 with HTTP Streamable endpoint at /mcp. It uses detailed logging to show exactly which tools are being called and their results.

Bridge Sessions (Record & Playback)

The Bridge Sessions feature allows you to record yours or an AI's tool calls and replay them later. This is useful for macros, versioning your creations, or setting up complex scenes consistently.

Recording a Session

To record all tool calls made to the bridge while the server is running:

uv run python -m src.main serve --record path/to/session.json --name "My Project" --description "Optional description"

Any tool calls made by n8n or other clients will be automatically saved to the JSON file.

Replaying a Session

To playback a previously recorded session:

# Default (Stateful - HTTP Streamable) - Recommended for speed
uv run python -m src.main play path/to/session.json

# Stateless mode (Standard HTTP) - Slower due to handshake overhead
uv run python -m src.main play path/to/session.json --transport stateless

Tip

Performance Note: Stateful mode is significantly faster for playback because it maintains a persistent connection. Stateless mode requires a full MCP handshake (Initialize/Discover) for every individual tool call in the recording, leading to noticeable overhead.

Session Format

Sessions are stored as JSON files containing metadata (name, description) and a list of command objects (tool name, arguments). Studio sessions may also include optional parameters (global, parametric values referenced from command arguments via ${name} tokens, including arithmetic expressions) and branches (named, jumpable subsets of the command list). Plain linear sessions with neither key work exactly as before.

Studio (Visual Editor)

Studio is the built-in visual editor for inspecting, editing, and replaying your recordings — a Vite + React app that talks directly to the MCP Bridge Server.

Setup (run from the studio/ directory, not the project root — it's a separate npm package from any Node tooling elsewhere in the repo):

cd studio
npm install

Development — starts a hot-reloading dev server (default http://localhost:5173):

npm run dev

Production build — outputs a static bundle to studio/dist/:

npm run build

Preview a production build — serves the dist/ output locally to sanity-check it before deploying:

npm run preview

Once built, the Bridge Server also serves Studio directly at http://localhost:8008/studio/ (no separate npm run dev needed) — if studio/dist/ doesn't exist yet, that route returns a short message with the build command instead of failing.

Studio talks directly to the MCP Bridge Server over HTTP (same /mcp endpoint used by n8n), so the Bridge Server must be running (uv run python -m src.main serve) for Studio to connect. It auto-detects the bridge on localhost:8008 then localhost:8000.

In Studio you can:

Edit session metadata (Name, Description, AI Model).
Filter commands by tool name.
Add & Edit Commands: an interactive modal with schema validation to discover tools, safely modify arguments, or add entirely new steps.
Edit tool arguments directly in the JSON editor cards, or toggle to a whole-session Guided / JSON view for bulk edits with a command outline for quick navigation. JSON mode includes live parametric-expression linting (undefined ${param} references, divide-by-zero, and syntax errors are flagged inline and block Apply).
Reorder or delete commands, and replay them from the ActionBar (Play All / Play / Play to Active / Stop / Undo / Redo / Reset / Clear Scene), always visible at the top of the Command Timeline dock regardless of whether the timeline's diagram is expanded or collapsed.
Define global parameters (${name} tokens, including arithmetic expressions like ${box_height} / 2) so command arguments stay reusable across resizes/edits.
Define branches — named, non-contiguous ranges of the command list you can jump between and run independently (e.g. a "plain" feature vs. a "with extra part" feature sharing a common base) — via the Branches panel's click-to-pick range builder, then pick and run one from the ActionBar's branch dropdown.
Command Timeline: a horizontal diagram of command order and branch ranges at the bottom of the window (resizable, click a tick to jump to that command). Long branch ranges can be collapsed into a single condensed marker (click a branch's range bar, Excel-style column grouping) to keep the diagram readable in sessions with many commands.
Export JSON to save your changes to a new file.

AI Assistant Panel (n8n-free Alternative)

Studio also includes a standalone AI Assistant — a resizable chat drawer (toggle with the header's 🤖 button) that lets you drive Blender with natural language, without setting up n8n at all. It talks directly to an LLM provider through the Bridge Server's /assistant/* endpoints, using the same MCP tool set as everything else in this project.

Multi-provider: switch between Anthropic (Claude), Google (Gemini), and OpenRouter (which unlocks most open models, including free ones) via a dropdown. Custom model IDs can be saved per provider and persist across restarts.
API keys: read from the Bridge's environment by default; a key pasted into the panel (stored only in your browser) overrides it — useful when the bridge host has no key configured, or when experimenting with a different provider.
Load STL: upload a model file directly from the panel; it's saved under BLENDER_ASSETS_DIR/uploads/ and imported into the scene automatically.
Vision-verified agent loop: when a tool call produces a screenshot or render, the image is attached to the conversation (in each provider's native image format), so the model can actually see the result of its own edits — catching floating or misaligned geometry that numeric verification alone would miss. Models without vision support get a graceful fallback to numeric-only verification.
Recordable: like every other client, assistant-driven tool calls flow through the same session recorder (serve --record), so an AI conversation produces a replayable session.json.

The community Benchy sail rig was designed almost entirely through this panel — including the print-design fixes (diamond-profile spars, flush bar tips) discovered by looking at real slicer output.

Demo Mode (No Bridge Required)

Live demo on GitHub Pages — auto-deployed by .github/workflows/deploy-studio.yml on every push to main that touches studio/.

Note

The connection check always runs in your browser, against localhost — never against the page author's machine. Studio's connection probe (studio/src/lib/api.js) always targets http://localhost:8008 (then :8000), regardless of where the Studio page itself is hosted. So:

If you have your own Bridge Server running locally, the hosted Pages demo will genuinely connect to it and drive your own Blender instance — hosting the UI once on Pages doesn't require running it locally too.
If you don't have a bridge running, the header shows "Offline" and a "▶ Try Demo Mode" button appears. Demo mode is always opt-in, never automatic — a genuinely offline bridge is never silently mistaken for a working connection.

Either way, no one visiting the Pages demo can ever reach another visitor's Blender instance — the browser only ever talks to localhost on the machine it's running on.

Either way, once in Demo Mode:

Loads a static tool catalog (studio/src/lib/demoTools.js) so the guided forms, schema validation, and the + New Command / Edit Command modals all work normally.
Simulates tool execution and playback (studio/src/lib/demoApi.js) instead of calling a real bridge — good for exploring the session editor, branches, and the Command Timeline with any of the community sample sessions.

What demo mode doesn't do: it can't talk to Blender, so nothing is actually rendered or exported, and the AI Assistant panel is disabled while it's active (it needs a live bridge and a real model API to do anything meaningful). Exit demo mode any time with the header's "✕ Exit Demo" button.

3. Configure n8n Workflow

Add MCP Client Tool node
Configure:
- HTTP Streamable Endpoint: http://localhost:8008/mcp
- Authentication: None
- Tools to Include: All
Connect to an AI Agent node

4. Development: Updating & Applying Changes

If you modify the addon code or the MCP server logic, follow these steps to ensure changes are applied:

Reload Scripts: In Blender, press F3 and type "Reload Scripts" (or use the shortcut Alt + R if configured).
Restart Blender Server: In the N-Panel, click Stop MCP Server and then Start MCP Server again.
Restart Python Server: Stop and restart the server with uv run python -m src.main serve.

Important

All Blender operations now run on the main thread via a command queue, ensuring stability and preventing dependency graph errors.

Available Tools

The server exposes 93 Blender tools — the full reference with descriptions and parameters lives in docs/tools.md (auto-generated from the tool schemas; regenerate with uv run python scripts/gen_tools_doc.py, which also fails loudly if the bridge schemas and the addon dispatch table ever drift apart).

Category	Tools	Highlights
Modeling	47	primitives, booleans, modifiers, `create_curve` 2D drafting (lines + true arcs), `create_watertight_plate`, room shells
Materials	7	PBR materials, hex colors, texture assignment
Collections	7	create/move/organize collections
Sculpting	7	brush-based sculpt strokes and remeshing
3D-Print Preparation	6	`check_mesh_for_printing`, `repair_mesh`, voxel remesh, import STL/OBJ/FBX, export STL/3MF
Scene & Diagnostics	4	scene/object info, viewport screenshots, distances
Animation	4	keyframes, playback range
Lighting & World	3	lights, light editing, HDRI/sky/color world background
Camera	3	cameras, look-at, active camera
Rendering	3	engine setup, frame renders
History / Undo	2	undo / redo

Example Usage in n8n

Example Interaction

User: "Create a red metallic sphere at position [0, 0, 2]"

Assistant: Perfect! I've successfully created a red metallic sphere at position [0, 0, 2]. Here's what was done:

Created a sphere named "RedMetallicSphere" with a radius of 1 unit at coordinates [0, 0, 2]

Applied a red metallic material using a polished metal preset with a red base color (#FF0000)

The sphere is now ready in your Blender scene with a shiny red metallic appearance.

User: "Duplicate the red sphere, move it to [5, 0, 2], and change its color to blue"

Assistant: Perfect! I've successfully completed all the tasks:

Duplicated the red sphere and created a new object named "BlueMetallicSphere"

Moved it to position [5, 0, 2]

Changed the material to blue metallic with a polished metal finish

You now have two metallic spheres in your scene: RedMetallicSphere at [0, 0, 2] (red) and BlueMetallicSphere at [5, 0, 2] (blue).

User: "Select both the original red sphere and the duplicate, then add a Subdivision modifier to both"

Assistant: Excellent! I've successfully:

Selected both spheres (RedMetallicSphere and BlueMetallicSphere)

Added a Subdivision Surface modifier (SUBSURF) to both objects with 2 levels of subdivision

This modifier will smooth and refine the geometry of both spheres, giving them a more polished and rounded appearance.

User: "Add a torus with minor_radius 0.1 and major_radius 25 at [2, 0, 0]"

Assistant: Perfect! I've successfully created a torus with the specifications you requested:

Name: Torus

Position: [2, 0, 0]

Major radius: 25 (distance from center to center of the tube)

Minor radius: 0.1 (thickness of the tube)

This creates a very thin, large torus - essentially a large ring with a thin profile. It's now in your Blender scene alongside the two metallic spheres!

Community Showcase

This project is powered by its community! Explore recorded sessions and documentation created by users to see what's possible with the Blender MCP:

Community Gallery: Browse all user-submitted projects and learn how to contribute your own recordings.
Condominium Tower: A complete guide to creating a procedural 20-story building with glass facade and balconies.
Boolean Pavilion: Demonstrates boolean operations, unified structures, and advanced lighting/camera setup.

Tip

Share Your Work: Have you built something cool? Check out our Contribution Guide to learn how to record, clean, and share your session with the community!

⚡ POWER TIPS: Avoiding Rate Limits

To prevent n8n or LLM "Too many requests" errors, follow these Stateless Power Working rules:

1. Avoid Selection-Based Workflows

❌ Slow (4+ turns): select_by_pattern('Wall_*') → create_material('M_Gray') → assign_material() ✅ Fast (1 turn): create_material(name='M_Gray', pattern='Wall_*')

2. Group by Collections

❌ Unreliable: Selecting individual objects. ✅ Reliable: create_material(name='M_Glass', collection='Cutters')

3. Bulk creation

If you need 10 objects, don't create them one-by-one. Use create_and_array or duplicate_object with count.

Architecture & Technical Design

This project uses a modular src/ structure to ensure maintainability:

graph TD
    A[main.py] --> B[server.py]
    B --> C[tools/ package]
    C --> D[modeling/ package]
    C --> E[scene.py]
    C --> F[materials.py]
    B --> G[connection.py]
    B --> I[sessions.py]
    B --> J[assistant.py]
    G --> H[Blender]

Transport Model

Although the MCP specification supports persistent HTTP Streamable sessions, many clients (including n8n) currently operate in a stateless execution model, performing:

Initialize → Discover Tools → Call Tool → Close

for each interaction.

The server uses the official HTTP Streamable transport introduced in MCP SDK 1.8.0+, which supports both stateful sessions and stateless requests.

The Stateless Fallback Mechanism

To ensure reliability across clients, the server implements a robust fallback strategy:

Protocol Resilience: If no active HTTP Streamable session exists, the server transparently handles standard JSON-RPC requests over HTTP.
Execution Isolation: Each tool call is processed independently, preventing session corruption or deadlocks.
Visual Success Indicators: Tool responses are prefixed with ✓ when successful. This helps the AI Agent’s conversational memory confirm task completion and avoid unintended re-execution loops.
Clear State Boundaries: Persistent state is intentionally separated:
- 🧠 Conversation memory → AI Agent (n8n Simple Memory)
- 🧩 Scene state → Blender runtime
- 🚀 MCP server → Stateless execution bridge

Architecture Diagram

n8n AI Agent 
      ↓
MCP Client (HTTP Streamable / JSON-RPC)
      ↓
MCP Server (ASGI)
      ↓
TCP Socket Bridge
      ↓
Blender Addon (Main Thread Queue)
      ↓
Blender Scene (Persistent State)

Testing

We use an integrated test suite to verify Blender tools and layout scenarios.

# Run the Arch layout test
uv run python tests/run_integration.py run --scenario arch

# Run the standard functional grid test
uv run python tests/run_integration.py run --scenario grid

# Run the 3D Printing tool verification test
uv run python tests/run_integration.py run --scenario print

# Run the Filament Name Tag & Stand generation test
uv run python tests/run_integration.py run --scenario filament_tag

Scenarios

Scenario	Key	Description
Grid Layout	`grid`	Functional grid test covering all tool categories
Arch Layout	`arch`	Architectural scene generation test
Print Validation	`print`	3D printing workflow: units, mesh repair, export
Filament Tag	`filament_tag`	Generates a 4-piece modular filament name tag & clip system (NameTagCard, AMSClip, StickonHolder, DeskStand)

See the Integration Testing Guide for full details on verification and benchmarking.

Code Quality & Standards

We enforce code quality standards using Ruff and Mypy. These are run automatically on GitHub Actions CI.

To run these checks locally:

# 1. Format code (strict black-compatible formatting)
uv run ruff format src/ tests/ blender_mcp_addon/

# 2. Run linter and check code complexity (McCabe <= 12)
uv run ruff check src/ tests/ blender_mcp_addon/

# 3. Auto-fix standard lint issues
uv run ruff check src/ tests/ blender_mcp_addon/ --fix

# 4. Run static type checking
uv run mypy src/

Troubleshooting

Server won't start: Install dependencies with uv sync

Connection failed: Ensure Blender MCP addon is running on port 8888.

Dependency Graph Error: If you see this, ensure you have the latest blender_mcp_addon package which implements the main-thread command queue.

Tools not appearing in n8n: Check the HTTP Streamable endpoint URL is correct (http://localhost:8008/mcp)

Acknowledgments

This project was inspired by blender-mcp by [ahujasid], which demonstrated the potential of MCP servers for Blender automation.

License

MIT License - See LICENSE file for details

Name		Name	Last commit message	Last commit date
Latest commit History 23 Commits
.github/workflows		.github/workflows
blender_mcp_addon		blender_mcp_addon
community		community
docs		docs
scripts		scripts
src		src
studio		studio
tests		tests
.env.example		.env.example
.gitignore		.gitignore
LICENSE		LICENSE
README.md		README.md
pyproject.toml		pyproject.toml
uv.lock		uv.lock

Folders and files

Latest commit

History

Repository files navigation

Blender MCP Server for n8n

System Architecture

Quick Start

1. Install Dependencies

Configuration

Installation

Method 1: Zip & Install (Recommended)

Method 2: Manual Copy (Developer)

Why a folder instead of a single file?

Usage

1. Start Blender MCP Addon

2. Start the MCP Bridge Server

Bridge Sessions (Record & Playback)

Recording a Session

Replaying a Session

Session Format

Studio (Visual Editor)

AI Assistant Panel (n8n-free Alternative)

Demo Mode (No Bridge Required)

3. Configure n8n Workflow

4. Development: Updating & Applying Changes

Available Tools

Example Usage in n8n

Example Interaction

Community Showcase

⚡ POWER TIPS: Avoiding Rate Limits

1. Avoid Selection-Based Workflows

2. Group by Collections

3. Bulk creation

Architecture & Technical Design

Transport Model

The Stateless Fallback Mechanism

Architecture Diagram

Testing

Scenarios

Code Quality & Standards

Troubleshooting

Acknowledgments

License

About

Topics

Resources

License

Uh oh!

Stars

Watchers

Forks

Releases

Packages 0

Uh oh!

Contributors 1

Languages

Packages