Blender MCP Server

Give an AI agent hands-on control of Blender: build geometry, apply materials, render, and check that a mesh is printable — 98 tools over the Model Context Protocol.

Works with Claude Desktop, Claude Code, n8n, or any MCP client.

uvx blender-mcp-bridge serve

That is the server. It drives a small addon inside Blender, which does the actual work — see Installation, a one-time step.

System Architecture

Two components, deliberately separate:

1. Blender MCP Addon: a plugin installed inside Blender. The local execution engine — it receives commands and manipulates the 3D scene.
2. MCP Bridge Server: a standalone Python server (blender_mcp_bridge/) that acts as the gateway. MCP clients connect to the Bridge, which forwards commands on to the running Blender addon.

graph LR
    Client[Claude / n8n / any MCP client] -- "MCP (HTTP Streamable)" --> Bridge[MCP Bridge Server]
    Bridge -- "Local TCP socket" --> Addon[Blender MCP Addon]
    Addon -- "Python API" --> Blender[Blender Engine]

Design rules for printable geometry

data/design-rules.json carries 3D-printing design rules — overhang limits, wall minimums, clearances — extracted from video sources by the sibling print-kb project, each citing the video and second it came from. It is plain self-contained JSON: no database, no dependency on that project at runtime.

See docs/design-rules.md for the schema, how to read it, and the MCP tools worth building on top.

Quick Start

1. Install the Blender addon

The addon is what actually touches Blender, so install it first — full steps under Installation. In Blender: Edit → Preferences → Add-ons → Install, pick the zipped blender_mcp_addon folder, enable it, then press Start MCP Server in the sidebar (N → Blender MCP).

2. Run the bridge

uvx blender-mcp-bridge serve

No clone, no virtualenv. It listens on http://127.0.0.1:8008/mcp/.

To hack on it instead, clone the repo and use uv:

uv sync
uv run blender-mcp-bridge serve

3. Point a client at it

This server speaks HTTP streamable, not stdio, so clients connect by URL rather than by spawning a command.

Claude Code — one command:

claude mcp add --transport http blender http://127.0.0.1:8008/mcp/

Claude Desktop — in claude_desktop_config.json:

{
  "mcpServers": {
    "blender": {
      "type": "streamable-http",
      "url": "http://127.0.0.1:8008/mcp/"
    }
  }
}

n8n — add an MCP Client node pointing at the same URL.

Then ask your agent to "create a 20 mm cube and check whether it is printable". It should appear in the Blender viewport.

Configuration

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

Only needed for the AI Assistant panel — each provider is offered in the UI only if one of its keys is set here (or pasted into the panel at runtime):

Installation

Method 1: Zip & Install (Recommended)

1. Zip the blender_mcp_addon folder (into blender_mcp_addon.zip).
2. Open Blender.
3. 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.
4. Search for "Blender MCP" and enable the checkbox.

Method 2: Manual Copy (Developer)

1. Copy the blender_mcp_addon folder to your Blender addons directory (substitute your actual Blender version for <version>, e.g. 4.2 or 5.0):
   • Windows: %USERPROFILE%\AppData\Roaming\Blender Foundation\Blender\<version>\scripts\addons
   • macOS: ~/Library/Application Support/Blender/<version>/scripts/addons
2. Restart Blender.
3. Enable "Blender MCP" in Preferences.

[!NOTE] bl_info declares a minimum of Blender 4.0, but STL import/export use the wm.stl_import / wm.stl_export operators introduced in 4.2 — so the 3D-print tools need 4.2 or newer. Developed against 5.0.

Why a folder instead of a single file?

As the addon grows, a single-file addon becomes unmaintainable — it is now ~6,400 lines across 22 modules. We've split the logic into functional packages (modeling, materials, anim, etc.) to make it professional, readable, and easier to extend.

Usage

1. Start Blender MCP Addon

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

2. Start the MCP Bridge Server

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

# Recording mode (Save all commands to a file)
uv run python -m blender_mcp_bridge.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 blender_mcp_bridge.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 blender_mcp_bridge.main play path/to/session.json

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

# Play a single branch (see Session Format below)
uv run python -m blender_mcp_bridge.main play path/to/session.json --branch sail_rig

# Override session parameters without editing the file (repeatable)
uv run python -m blender_mcp_bridge.main play path/to/session.json --param box_height=40 --param wall=2.4

[!NOTE] If a session defines branches and --branch is omitted, the first branch plays automatically rather than the whole command list. Any parameter not overridden by --param falls back to the session's own default from its parameters block.

[!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

Both dev and build first run npm run prebuild-assets, which regenerates two sets of static assets — you don't invoke these directly:

The copies are generated rather than committed, so they can never drift from the community/ source of truth. Add a new community folder with a session.json and it shows up in the dropdown on the next npm run dev/build — no manual step.

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 blender_mcp_bridge.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.

In all cases, 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.

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

1. Add MCP Client Tool node
2. Configure:
   • HTTP Streamable Endpoint: http://localhost:8008/mcp
   • Authentication: None
   • Tools to Include: All
3. 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:

1. Reload Scripts: In Blender, press F3 and type "Reload Scripts" (or use the shortcut Alt + R if configured).
2. Restart Blender Server: In the N-Panel, click Stop MCP Server and then Start MCP Server again.
3. Restart Python Server: Stop and restart the server with uv run python -m blender_mcp_bridge.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 98 tools — 95 forwarded to Blender, plus 3 answered locally from the design-rule export. 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).

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

Real parts, modelled by an agent driving this bridge. Two are in this repo so you can try one immediately:

blender-mcp-bridge play community/cable_comb/session.json

The full gallery — the benchy, architectural scenes, the modular profile rack, threaded bolts, the Y-zipper teardown and more — lives in its own repository, so installing the bridge does not mean downloading a print archive:

blender-mcp-community

[!TIP] Share your work: recorded something good? The contribution guide covers how to record, clean and submit a session.

⚡ 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 the modular blender_mcp_bridge/ package 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]
    I --> K[params.py]
    A --> L[config.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:

1. Protocol Resilience: If no active HTTP Streamable session exists, the server transparently handles standard JSON-RPC requests over HTTP.
2. Execution Isolation: Each tool call is processed independently, preventing session corruption or deadlocks.
3. 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.
4. 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

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

Unit Tests

The scenarios above need a live Blender. Two pure-Python suites don't — they cover the session record/replay round-trip and schema hygiene across all 94 tool definitions:

uv run --with pytest python -m pytest tests/ -q

See the Unit Testing Guide. Note these are not run by CI, which covers only ruff and mypy.

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 blender_mcp_bridge/ tests/ blender_mcp_addon/

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

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

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

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