There are three ways to connect AI to Blender: an MCP server, a Python script, or a generator that makes 3D models from text or an image. I use the middle one: Claude Opus 5.5 builds the world of my own work-in-progress game Noční linka (Night Line) as scripts that I run in Blender without a window. I describe MCP and the generators from official pages and press coverage; I have no first-hand experience with them.
Blender MCP lets Claude control an open Blender. The official add-on (Blender Lab, version 1.0.3) requires Blender 5.1 or newer and runs model-written code without safeguards, so Blender recommends a virtual machine. A Python script (bpy) is a readable source of the model and runs from the command line; that is how my game models are made. A 3D generator from an image (Meshy, Rodin) is quick for an isolated object, but the output needs retopology and licences differ. Blender itself has no generative AI and none is planned.
- Scripts without MCP. I have Claude Opus 5.5 write the 3D models for my game Noční linka (alpha, my own project, how it is made) as Python scripts for Blender. Blender 4.1.1 runs windowless from the command line, with Cycles rendering on the CPU.
- Independent review. GPT-6 Astra audited the models from renders: on the village houses it criticised bare sides and flat doors, and on the train and the mill repeated stains and plaster outlines.
- Animation fixes in Blender. The mop and the pole grip on the bus passed (palm anchor deviation of at most 4.54 mm and 2.19 mm, measured on bones), and the coat stayed HOLD after three rounds.
- What I have not tried: Blender MCP or a 3D generator, and I did not compare Astra with Opus on the same task.
What are the three ways to connect AI to Blender?
| Route | How it works | Good for | Risks |
|---|---|---|---|
Blender MCP |
Add-on and MCP server: Claude edits the open scene |
Scenes, lights, materials, batch edits |
Code runs without safeguards; fails on organic shapes and animation |
Python script (bpy) |
The model writes a script, Blender runs it without a window |
Repeatable models, fixing one detail, renders |
Someone has to review the result |
Generator from text or image |
You upload an image, the service returns a 3D model |
An isolated object, a concept |
Dense mesh, retopology, licences |
MCP and scripts both use the same Blender Python API; they differ in who runs the code and how.
What is Blender MCP and how do you connect it to Claude?
MCP (Model Context Protocol) is an open standard for connecting AI applications to tools; I explain it in the article on agents and the harness. Blender has two distinct implementations.
The official server from Blender Lab. According to Blender's page, it is an add-on, version 1.0.3, for Blender 5.1 and newer (5.1 came out on 17 March 2026, 5.2 LTS on 14 July 2026). You install three parts: the add-on, a client and the MCP server. The page warns that model-written code runs without safeguards against deleting or sending out data, and recommends a virtual machine or a computer without sensitive data. It names only the llama.cpp client, and Claude is not mentioned on the page. In its announcement of 28 April 2026 Anthropic writes that the Blender developers built an MCP connector that is now officially available for Claude. If the connector "will not install", check that your client supports .mcpb bundles; according to the page, llama.cpp does not yet.
The community server ahujasid/blender-mcp. The GitHub project (MIT licence, about 31.8 thousand stars on 20 October 2026) was not made by the Blender Foundation. According to its README it works with Claude Desktop, Claude Code, Cursor, Codex and VS Code, and it reaches Poly Haven (CC0), Sketchfab and the Rodin, Hunyuan3D, Meshy and Tripo generators. The execute_blender_code tool runs arbitrary Python, the socket has no authentication or encryption and should stay on localhost. An optional safe mode checks scripts before they run.
What works in practice. According to a press summary, not my own measurement, MCP reliably handles scenes, lights, materials and batch scripts. It fails at organic modelling, topology, Geometry Nodes, animation and anything that needs visual judgment, such as proportions.
On 1 May 2026 the Blender Foundation stated that no generative AI feature exists or is planned in Blender. Anthropic supported Blender financially; after community feedback the foundation accepted it as a one-time donation, not as Development Fund membership (source). So MCP is an external bridge: the model runs outside Blender and uses its Python API.
What is Reddit saying about Blender and AI?
Reddit blocks automated reading and I did not open the individual threads, so I do not take any claim from it as fact. According to my notes, on 20 October 2026 I only skimmed Google result snippets for the queries site:reddit.com blender mcp, site:reddit.com claude blender and site:reddit.com blender cli python ai. I do not list links to specific threads, comment counts or the content of replies here, because they cannot be verified.
Only three questions recurred in the snippets: whether MCP or a Python script is better, which model does better in Blender (Claude Opus 5.5 or GPT-6 Astra), and how safe it is to run model-written code. That is an impression from titles, not a measurement, and I build nothing on it. My own experience is only with scripts; about MCP and the difference between the routes I know only from official pages and press coverage. For security, see the warning below.
How I do it: Python scripts without MCP
Noční linka is my own game in Godot 4.7.1 (more in AI and game development), still an alpha and unreleased, with third-party assets only under CC0. The 3D models of its world, meaning panel blocks, village houses, a train, cars, a bus, street furniture, a mill and a forester's lodge, were created by Claude Opus 5.5 as Python scripts for Blender (commits from 7 and 9 October 2026). I run them in Blender 4.1.1 from the command line, without a window, and Cycles renders on the CPU. No MCP, no generator. According to Blender's page the official add-on requires Blender 5.1 or newer, and I run 4.1.1.
The main advantage: the source of a model is a readable script. When one thing is wrong on a village house, I fix that detail and do not regenerate the whole scene. I export the finished model to GLB, load it again in Godot and check it.
A script cannot verify itself, though. The author, Opus or GPT-6 Astra, is therefore paired with an independent critic who judges the actual renders, not the author's claims; I describe the eight production gates in the article on AI and game development. On 10 and 11 October 2026 Astra audited the models from Opus. For the bus it praised the readable silhouette, frames, handrails and seats. For the village houses it criticised bare sides and backs, flat doors and overly regular roofs that look like a building kit up close. On the panel blocks the dirt map repeats, the train has the same spots on the roof and the mill has repeated outlines of damaged plaster.
I did not run a direct Astra versus Opus comparison on the same task, so I do not claim that one is better. I did not compare scripts with MCP either, I only describe my own workflow.
The website videos are made the same way. The service animations on le-artist.cz are three.js scenes, but selected videos are rendered in Blender Cycles from scripts in the motion/blender folder; captions are added separately in Remotion, according to a comment in my script for the marketing video. The Cycles video, with Czech captions, plays on the Czech page Online marketing ("Jak vedu zákazníka od hledání k poptávce", 18 s), and likewise on the Czech pages for web development and e-shop redesign. I render on the CPU with a single Blender instance: two at once crashed my AMD graphics driver and froze the computer (a note from 2 October 2026 in my render script).
WebGL computes the image in the browser in real time; Cycles renders it frame by frame with more physical accuracy, including reflections and soft shadows, but only as a finished video.
Which AI generator should you pick for a 3D model from an image?
I did not use a generator in the game and I tested none. The data comes from official pages and repositories, verified on 20 October 2026; I opened the Meshy and Rodin price lists again on 22 October 2026. Prices change, so check them before you buy.
| Tool | Price | Where it runs | Licence and caveats |
|---|---|---|---|
Meshy (7.1) |
Free 100 credits; per the price list in Czech koruna on 22 October 2026 Pro 420 CZK/month (1,000 credits), Premium 840 CZK (3,000), Ultra 2,100 CZK (8,000); beside the amounts the page gives both "billed monthly" and "billed yearly" and discounts the first month, so confirm the exact price at checkout |
Cloud |
Free: CC BY 4.0; paid: private assets you own |
Creator $30/month, Business $120/month when billed monthly (on 22 October 2026; $24 and $96 a month when billed yearly; the price list shows the $ sign without naming the currency) |
Cloud |
Paid plans: unlimited export and any use, per the price list |
|
TRELLIS.2 (Microsoft) |
No fee (MIT), 4B parameters |
Your own machine: NVIDIA GPU with 24 GB, Linux only per the repository |
PBR GLB output; some dependencies have their own licences |
Hunyuan3D 2.1 (Tencent) |
No download fee |
Your own machine: 29 GB VRAM for shape and texture |
The licence explicitly excludes the EU, the United Kingdom and South Korea |
Tripo |
I have not verified the price |
Not verified |
Not verified |
Meshy 7 became generally available on 10 August 2026, and the rollout of version 7.1 began on 10 September 2026 (per the Meshy blog, checked on 22 October 2026). For a Czech studio the licence of Hunyuan3D 2.1 is decisive, because it does not apply in the EU.
Generator output is often a dense triangle mesh, so for a game or animation it needs retopology, for example in QuadriFlow or Instant Meshes. For my game I wrote myself a rule: a generator at most as a candidate for an isolated prop, and only after the licence is checked. I do not want the whole depot as one generated mesh, because I need to control passages, collisions and individual parts. That is a note from my 5 October 2026 project notes, so a plan, not a test result. You can prepare the input image with the overview of AI image generators.
When should you use MCP, a script or a generator?
- Quickly editing an open scene (lights, materials, export): Blender MCP, ideally in a virtual machine and with your work saved.
- A repeatable, reviewable model (game, catalogue, video): a script run without a window, the source in version control, renders reviewed by someone other than the author.
- An isolated object from an image (prop, background): a generator, with retopology and a checked licence.
- An organic shape or animation: do not rely blindly on any route, MCP reportedly fails here.
- Deciding whether it looks good: that belongs to a person or an independent critic looking at the actual render, not to the model. There is an example in the article on GPT-6 Astra.
Is it safe, and what about licences?
The official add-on and the community server both run model-written code in Blender. Work in a virtual machine or on a computer without sensitive data, and save before every run. The same applies to scripts.
Read the licence of each tool separately. Meshy Free outputs are under CC BY 4.0, Hunyuan3D 2.1 does not apply in the EU, and for TRELLIS.2 some dependencies have their own terms. I discuss Steam's rules for AI content in the article on AI and game development.
Need a finished video for a campaign rather than tool experiments? I make promo videos, and you can see a sample of a Cycles render on the Czech Online marketing page. If you want to work with AI safely in a company, from a process audit to operation, that is AI automation.
FAQ
What does Reddit say about Blender and AI?
I cannot document that: I did not open the threads. According to my notes from Google search snippets of 20 October 2026, three topics recurred: whether to use MCP or Python scripts, how Opus 5.5 and GPT-6 Astra do in Blender, and how safe it is to run model-written code. That is an impression from titles, not a measurement, and I build nothing on it.
Does Blender have built-in AI?
No. On 1 May 2026 the Blender Foundation stated that no generative AI feature is available or planned in Blender. AI connects from outside: through an MCP server, Python scripts or external model generators.
How do you connect Claude to Blender?
Either through the official Blender Lab add-on (version 1.0.3, Blender 5.1 or newer; it does not name Claude) with the MCP connector that, according to Anthropic on 28 April 2026, the Blender developers built, or through the community server ahujasid/blender-mcp, which according to its README works with Claude Desktop and Claude Code. You install the add-on in Blender and add the server to your client. Without MCP, you can simply have Claude write scripts and run them from the command line.
Is Blender MCP safe?
Not without precautions. The official add-on runs model-written code without safeguards and Blender recommends a virtual machine. The community server runs arbitrary Python and its socket has no authentication or encryption, so keep it on localhost. The safe mode is optional.
What is the best AI generator for a 3D model from an image?
I have not tried them, so I will not rank them. Meshy is the simplest cloud route (100 free credits, paid plans from 420 CZK a month per the price list of 22 October 2026; confirm the exact price at checkout). TRELLIS.2 has no fee but needs an NVIDIA GPU with 24 GB. Hunyuan3D 2.1 has a licence that excludes the EU. Plan for retopology of every output.
Can I use Meshy models commercially?
According to Meshy's price list (verified 20 and 22 October 2026) yes, but the terms depend on the plan. Free outputs are under CC BY 4.0, so with credit to Meshy. On paid plans the assets are private and you own them. Read the current terms; this is not legal advice.