mcp-blender-agent
Deterministic Blender control for AI agents, typed tools, not improvised Python.
接入你的智能体
從發布者安裝。不是 PlugMCP 端點。
遠端 URL 還是本機指令? 兩種安裝方式的差異
The point nobody discusses in Blender MCP demos: the problem isn't the AI. Most MCPs just delegate Python scripts to Blender, each run generates different code, no grouped undo, no output schema. When it fails you don't know if it was the model, the script, or the scene state.
You solve this outside the addon. Each operation becomes a typed tool: input validated, output structured as { ok, data, refs, nextSteps } with IDs that chain into the next call. Composite mutations end with a single undo push. The agent stops improvising Python. Same input, same result. Ctrl+Z undoes the entire operation.
~216 tools across 12 domains, a separate tool for each Blender operation: modeling, rigging, animation, materials, shader/geometry nodes, rendering, FBX/glTF/USD/Alembic export with every exporter parameter exposed. Runs on Blender 4.2 LTS+ (5.x recommended).
→ Architecture: .claude/docs/ARCHITECTURE.md
You install the MCP by running one prompt in whatever agent you use, Claude Code, GitHub Copilot, Cursor, Codex, opencode, Claude Desktop. The agent detects which harness it's running in, configures itself, and walks you through the one manual step (installing BlenderAgent.zip in Blender's Add-ons UI). No clone, no build, the server runs via npx (prereq: Node 18+).
Paste this into your agent chat:
Install @pobruno/blender-agent into this workspace. Fetch and read
https://raw.githubusercontent.com/PoBruno/mcp-blender-agent/main/install/AGENT-INSTALL.md
and run every phase. Detect which agent harness you are running as (Claude Code,
GitHub Copilot, Cursor, Codex, opencode, or Claude Desktop) and follow the
config-target and skill-placement tables for that harness: merge the
`npx -y @pobruno/blender-agent@latest` server entry into my MCP config, copy the
bundled context skill (`npx -y @pobruno/blender-agent --print-skill-dir`) into my
harness, and inject the managed block from MANAGED-BLOCK.md into my primary
instruction file. Ask before anything destructive or global. Print the path from
`--print-addon-zip` so I can install it in Blender's Add-ons UI.
Per-harness prompts and workflow recipes: install/PROMPT-TEMPLATES.md · Manual reference: install/INSTALL.md
git clone https://github.com/PoBruno/mcp-blender-agent.git
cd mcp-blender-agent/Tools
npm install && npm run build
npm test # requires Blender on PATH or BLENDER_BIN
MIT. Not affiliated with the Blender Foundation.
MCP server giving AI agents complete control of Blender, modeling, rigging, animation, materials, render, export. Mirror architecture of mcp-unreal-agent.
工具
尚無工具清單。遠端伺服器會在首次健康檢查後取得工具清單。
常見問題
在本頁複製適用於 Cursor、Claude Desktop、Claude Code 或 VS Code 的 stdio 設定。它使用發布者的安裝方式,而不是 PlugMCP 代理伺服器。
這筆記錄未列出必要的環境變數。
MCP Blender Agent 透過 stdio 在本機執行。
上次檢查時間:2026-10-28。 MCP server giving AI agents complete control of Blender, modeling, rigging, animation, materials, render, export. Mirror architecture of mcp-unreal-agent.