Blender MCP
Introduction
A Blender MCP server is an MCP (Model Context Protocol) server that connects your local Blender instance to an AI client. The model calls named tools, and the server runs them in your running Blender session through the Python API.
The public ecosystem is built around modeling: ask the assistant to build a scene. For investigation work, the higher-value tools are the ones that read and organize a scene: analysis, naming, inventory, checklists, and verification. This page features three servers, selected for how well they serve that use:
- blender_mcp is the official Blender project server and the strongest fit for organizational work.
- blender-ai-mcp is verification-first: deterministic measurement and assertion tools.
- MCPBlender / blender-mcp is the most popular server (25,000+ stars), built for modeling and asset pipelines. Most people find it first. Its default-on telemetry is the caveat for real work.
- Cloud clients (Claude, Gemini): Scene metadata (object names, positions, materials) travels to the cloud provider. The 3D geometry stays local.
- Local LLMs (Ollama, llama.cpp): Everything stays on your machine.
The official server executes LLM-generated code without guards, and its documentation recommends a virtual machine for untrusted work. For sensitive cases, use a local model and a scratch copy of the file.
Key Takeaways
Organizational Value
The tools worth using are the read and organize tools: scene analysis, data block renaming, inventory queries, checklist verification, and documentation. The output is reviewable and belongs in a case record.
Three Featured Servers
The official Blender server for analysis and naming, PatrykIti's blender-ai-mcp for deterministic verification, and MCPBlender for modeling and asset pipelines. Each fits a different part of the workflow.
Verification Discipline
LLM analysis is not deterministic. Verify every result in the Outliner, diff the object list before and after, and keep a backup copy before each session.
Installation
Three criteria, applied strictly:
- Organizational value. Scene analysis, renaming, inventory, checklists, reporting, not just prompt-to-geometry.
- Privacy and security posture. No default-on telemetry, clear warnings about code execution, a path to local models.
- Maintenance. Recent commits, active issue handling. The landscape moves monthly. Stars measure attention, not fitness.
| Server | Maintainer | Focus | Privacy posture |
|---|---|---|---|
| blender_mcp | Blender project (official) | Scene analysis, data block renaming, documenting complex setups, checklists | No telemetry. Executes LLM code without guards, so use a VM for untrusted work |
| blender-ai-mcp | PatrykIti (Apache-2.0, active in 2026) | Verification-first: bounded tools, measurement and assertion, scope and relation graphs | Runs the server in Docker, which isolates it from your machine |
| MCPBlender / blender-mcp | Community (25,000+ stars) | Modeling, asset pipeline (Poly Haven, Sketchfab, Poly Pizza), code execution | Telemetry on by default, may train AI models. Disable it for real work |
The Official Blender Server
Installation
The official server needs three pieces: the add-on, an LLM client, and the MCP server.
- Add-on (v1.0.0, Blender 5.1+). Download
mcp-1.0.0.zipfrom the Blender Lab repository. Drag the zip onto the Blender interface to install, or use Install from Disk in Edit → Preferences → Add-ons. - LLM client. Blender has no built-in LLM connection. The documented path is llama.cpp; any MCP client works.
- MCP server. Use the
.mcpbbundle for newer clients, or install from source. See the Setup wiki.
Usage
Start with a read-only query· baseline
Ask for the scene structure before you ask for changes. This gives you a baseline to verify against later.
Run one organizational task at a time· diff in Outliner
Renaming, then inventory, then checklists. Each task is diff-able in the Outliner.
Verify every result· compare before/after
Inspect the Outliner, re-run the query manually, and compare the object list before and after.
Organizational prompts that work
These are tested examples from the Blender project's own page:
- Fix typos in data blocks. "Fix typos in my collection hierarchy: GRP-rocks should be GRP-pebbles, LGT-Lights should be LGT-lights."
- Apply naming conventions. "Rename the Camera to CAM-main, the key light to LGT-sun-key, the fill to LGT-sun-fill."
- Query data relationships. "Which objects are using the material pebbles?"
- Debug the scene. "What object has the highest polygon count? Ignore objects that are not linked to any scene."
- Analyze performance. Find objects with high polygon counts relative to the area they cover on screen, and suggest what to decimate or replace with a texture.
- Document a complex setup. "Explain what this Geometry Nodes tree does, add color-coded frames, and write the analysis into a Text data block."
- Run a checklist. "Check my scene: meshes must be manifold, all objects must have materials, naming must follow the convention, no absolute paths."
blender-ai-mcp (PatrykIti)
The verification-first alternative. Instead of generating ad-hoc Python for every request, the model calls tools with validated parameters against a stable API, in three layers: atomic tools, macro tools (task-sized operations), and workflow tools (bounded multi-step processes).
What makes it different for investigation work:
- Goal-first sessions.
router_set_goal(...)tells the system the intent before any action. - A truth layer.
scene_measure_*andscene_assert_*tools give deterministic inspection and assertion. The project's own words: vision assists interpretation, deterministic measurement provides the final truth layer. - Artifacts. Read-only scope graphs (anchor, core, accessory roles), relation graphs, and view diagnostics (frame coverage, occluded or off-frame verdicts per camera) that can be attached to a record.
- Isolation. The MCP server runs as a Docker image (
ghcr.io/patrykiti/blender-ai-mcp). The Blender add-on starts a local RPC server on port 8765.
Setup is heavier (Docker plus add-on), and the project is small (56 stars) but actively developed with CI and a large test suite. Use it when the deliverable is a verified result, not a generated scene.
Scene Organization & Analysis
The patterns that matter, regardless of server:
- Naming and conventions. Prefix conventions (CAM-, LGT-, GEO-) make scenes auditable and exportable.
- Inventory queries. Data questions in plain language: which objects use a material, which are unlinked, which is the heaviest.
- Checklist verification. A case checklist the model runs and reports against, with you verifying each failure.
- Performance analysis. Polygon count against screen area flags the objects that hurt viewport speed.
- Documentation. Complex setups explained into Text data blocks, with frames and color coding.
- Reports. Scene structure and statistics exported to support the investigation documentation.
Two caveats. LLM analysis is not deterministic. The official server's own example shows a first pass missing a Solidify modifier that doubled an object's poly count. And where a server offers deterministic measurement tools (blender-ai-mcp), prefer them over LLM-generated estimates.
How It Works
Three components, one trust boundary:
- The add-on runs inside Blender and opens a local socket. It executes whatever the server sends, as your user, inside your running session.
- The MCP server is a small process that translates the client's tool calls into commands for the add-on.
- The client is where the model runs. If the client is cloud, scene metadata travels to that cloud. If it is local (Ollama, llama.cpp), nothing leaves the machine.
The trust boundary is the socket. Whatever the model decides to call, the add-on executes with your permissions. That is the flexibility and the risk in one sentence.
In an Investigation
:::details Details
- Method: Document the MCP server (name + version), the client and model, and the Blender version. Paste the exact prompt you used for each operation.
- Decision Log: Record why an AI-assisted step was used for a given operation (for example: bulk renaming of 400 imported objects to the case naming convention).
- Verification: Inspect each change in the Outliner. Compare the object list before and after. Keep a backup copy of the .blend before every MCP session.
- Sensitive cases: Use a local model so scene data never leaves the machine. Work on a scratch copy. For untrusted content, run Blender in a virtual machine, as the official documentation recommends. :::
Limitations
- The server executes LLM-generated code without guards. The official documentation recommends a virtual machine for untrusted work.
- Prompt injection. The model reads content (model files, web pages, document text) that can carry hidden instructions, and a server with code execution will act on them. Prefer servers with named, validated operations over raw code execution, and review each tool call.
- Reliability. LLM analysis misses things (modifiers, hidden collections). Verify in the Outliner. Prefer deterministic measurement tools where available.
- Telemetry varies by server. MCPBlender collects usage by default and, with consent, prompts, code, scene metadata, screenshots, and action trajectories, data it may use to train AI models. The official and PatrykIti servers do not run a default telemetry pipeline.
- Local models are slower and weaker than cloud models. 14B+ class for complex tasks.
For more advanced scene management techniques, check out our Scene Management Guide.
Resources
- Blender MCP technique page (server selection, cloud and local setups, organizational automation)
Sources
- Blender Lab: Blender MCP Server (official server, setup, and tested organizational example prompts)
- Blender Lab repository: blender_mcp (source, releases, Setup wiki)
- GitHub: blender-ai-mcp (PatrykIti) (verification-first server, truth layer, Docker deployment)
- GitHub: MCPBlender (blender-mcp) (popular modeling/asset server, formerly ahujasid/blender-mcp; telemetry details in its README)
- TechDrawAI: CAD MCP Servers, What Actually Works in 2026 (ecosystem landscape, security guidance, star counts as of 8 Aug 2026)
- Model Context Protocol: spec