blender-mcp
An MCP server that builds and modifies Blender .blend files, supporting both headless generation and live interaction with an open Blender window.
README
blender-mcp
An MCP server that builds Blender .blend files. It reaches Blender two ways:
- Headless — every call starts a fresh
blender --backgroundprocess. Deterministic, always available, and the right choice for generating files. - Live — a small Blender addon opens a local port so the server can drive the Blender window you already have open, and you watch it happen.
Both routes run the same handler code, so a script behaves identically either way. A local web UI lets you upload reference images and run generation jobs with the same tools.
Requirements
| Blender | 4.2 or newer (developed against 5.1) |
| Python | 3.12+ for the server (Blender brings its own) |
| uv | dependency and venv management |
Setup
uv sync # create .venv and install
cp .env.example .env # then set BLENDER_MCP_EXECUTABLE
uv run python scripts/build_addon.py --install # build + install the Blender addon
If you skip .env, the server searches your PATH and the standard
Program Files/Blender Foundation/Blender <version> folders on every drive, preferring
the newest version it finds.
Turning on the live bridge
The addon is installed but idle until you start it:
- Open Blender.
- Press <kbd>N</kbd> in the 3D viewport to show the sidebar.
- Open the MCP tab and click Start MCP Bridge.
Tick Start listening on launch in the addon preferences to skip this every time.
Connecting an MCP client
.cursor/mcp.json already registers the server for this project. To make it available
everywhere, copy that block into ~/.cursor/mcp.json:
{
"mcpServers": {
"blender": {
"command": "uv",
"args": ["run", "--directory", "C:\\Users\\josef\\Projects\\blender-mcp", "blender-mcp"]
}
}
}
Quality workflow
Generation quality comes from looking, not from a single script dump:
list_references/view_reference— open guide images firstcreate_blendfromstudioorproduct- Build in small
run_scriptsteps;apply_materialfor metal/plastic/rubber/glass/wood preview_views— multi-angle EEVEE previews; compare to refs withview_image- Fix, preview again
- Final
render/render_and_view(CYCLES only for finals; EEVEE while iterating)
Local web UI
uv run blender-mcp-web
# open http://127.0.0.1:8765
Upload reference images, enter a prompt, and browse jobs under output/jobs/.
- With
BLENDER_MCP_LLM_API_KEYin.env— the site runs a bounded multimodal agent that calls the same Blender tools (OpenAI-compatible or Anthropic viaBLENDER_MCP_LLM_PROVIDER). - Without a key — uploads and job pages still work; open the job in Cursor and ask the
MCP agent to build from
output/jobs/<id>/refs.
Tools
| Tool | Purpose |
|---|---|
blender_status |
Install, output/refs dirs, templates, materials, live bridge |
list_references / view_reference / add_reference |
Guide image library under output/refs |
create_blend |
New file from a template + optional script |
run_script |
Arbitrary bpy Python |
apply_material |
Procedural presets: metal, plastic, rubber, glass, wood |
inspect_blend |
Scene graph / materials / cameras / lights |
preview_views |
Multi-angle EEVEE previews for critique |
render / render_and_view |
Still PNG; the latter returns the image to look at |
view_image / view_preview |
Load a PNG for vision |
viewport_screenshot |
Live viewport capture |
list_outputs |
Files under output/ |
open_in_live_session |
Open a .blend in the live Blender window |
Writing scripts
run_script and the script argument of create_blend execute in a namespace where
bpy, mathutils, math, D (bpy.data) and C (bpy.context) are already bound.
Send data back by assigning to result or calling emit(value).
Templates
| Name | What you get |
|---|---|
empty |
Nothing at all |
default |
Blender's startup scene: cube, camera, light |
studio |
Grey floor, three-point area lighting, 50mm camera, 1920×1080 |
product |
White seamless backdrop, soft lights, 85mm product framing |
Layout
addon/blender_mcp_addon/ Blender extension + shared handlers (incl. preview_views)
src/blender_mcp/
server.py MCP tools
refs.py reference image library
materials.py procedural material presets
templates.py starter scenes
agent/ bounded multimodal generation loop
web/ FastAPI UI (templates + static)
headless.py / bridge.py Blender process / live socket
scripts/ build, install, verification
output/ .blend files, renders, refs/, jobs/
Testing
uv run pytest -m "not blender" # fast
uv run pytest # includes real Blender runs
uv run python scripts/check_server.py
uv run python scripts/verify_live_bridge.py
Notes and limits
run_scriptexecutes arbitrary Python inside Blender. Keep the bridge on loopback.- Long renders block the live UI until they finish.
viewport_screenshotis live-only.- Cycles uses CPU unless you enable a GPU device in Blender preferences.
- The web agent needs a multimodal model that supports tool calls and images.
Licence
MIT
推荐服务器
Baidu Map
百度地图核心API现已全面兼容MCP协议,是国内首家兼容MCP协议的地图服务商。
Playwright MCP Server
一个模型上下文协议服务器,它使大型语言模型能够通过结构化的可访问性快照与网页进行交互,而无需视觉模型或屏幕截图。
Magic Component Platform (MCP)
一个由人工智能驱动的工具,可以从自然语言描述生成现代化的用户界面组件,并与流行的集成开发环境(IDE)集成,从而简化用户界面开发流程。
Audiense Insights MCP Server
通过模型上下文协议启用与 Audiense Insights 账户的交互,从而促进营销洞察和受众数据的提取和分析,包括人口统计信息、行为和影响者互动。
VeyraX
一个单一的 MCP 工具,连接你所有喜爱的工具:Gmail、日历以及其他 40 多个工具。
graphlit-mcp-server
模型上下文协议 (MCP) 服务器实现了 MCP 客户端与 Graphlit 服务之间的集成。 除了网络爬取之外,还可以将任何内容(从 Slack 到 Gmail 再到播客订阅源)导入到 Graphlit 项目中,然后从 MCP 客户端检索相关内容。
Kagi MCP Server
一个 MCP 服务器,集成了 Kagi 搜索功能和 Claude AI,使 Claude 能够在回答需要最新信息的问题时执行实时网络搜索。
e2b-mcp-server
使用 MCP 通过 e2b 运行代码。
Neon MCP Server
用于与 Neon 管理 API 和数据库交互的 MCP 服务器
Exa MCP Server
模型上下文协议(MCP)服务器允许像 Claude 这样的 AI 助手使用 Exa AI 搜索 API 进行网络搜索。这种设置允许 AI 模型以安全和受控的方式获取实时的网络信息。