What is blenderwright?
blenderwright, formerly blend-ai, from the GitHub organisation HoldMyBeer-gg, pairs a Python MCP server with a Blender extension so a model can build and edit scenes through named tools instead of arbitrary scripts. It registers 276 tools in 32 modules: objects, modifiers and mesh editing, 25 material and shader-node tools, lighting, cameras, 16 animation tools covering keyframes, drivers and NLA strips, render settings and baking, sculpting setup, UV mapping, physics, geometry nodes, armatures, Grease Pencil, the video sequencer and import and export in formats such as FBX, glTF, USD and STL. 13 prompts carry guidance on topology, real-world scale, lighting and PBR materials.
It is a separate project from MCP for Blender by ahujasid, which works mainly by running Python. blenderwright is MIT-licensed, has about 200 GitHub stars and released v3.1.0 on 26 October 2026. It targets Blender 5.2 LTS only and refuses to install on 5.1 or earlier.
How to install blenderwright
uvx blenderwrightIs blenderwright worth it?
blenderwright's argument is structure. Where MCP for Blender exposes nine tools and leans on execute_blender_code, this server gives the model 276 typed operations, each checked by validators for names, paths, numeric ranges and allowed values before it reaches Blender. A shader graph is built from a list of 93 permitted node types; a geometry-nodes tree can be built in a single call. The README's showcase renders, a saloon and a space-shuttle launch, each took one or two prompts and about twenty minutes, by the maintainer's account.
Security is better thought through than in most Blender servers. The add-on binds to 127.0.0.1 on port 9876, and because any local program can reach that port, every connection must pass a challenge-response handshake against a random token written to a file only your user can read. Raw Python is refused until you tick a per-session box in the N-panel, every snippet is echoed to the console, and 25 imports including os, subprocess and socket are blocked. The README is candid that this namespace is a guard against accidents, not a sandbox. Files the model opens or imports never auto-run their scripts.
The catch: breadth costs context. Loading all 276 schemas takes roughly 48,000 tokens per request, which suits Claude-class clients and swamps a local model with a 32,000-token window. The auto toolset starts with a core set and loads others on demand, but only in clients that honour MCP's tools/list_changed notification. Brush strokes in sculpting are not exposed, shader graphs are still built one node at a time, there is no grouped undo, and the server and add-on must be updated together because each refuses a mismatched protocol version.
Testing is unusually serious for a community project: more than 2,900 unit tests and over 600 live tests run every tool against a real Blender 5.2. The project is young, created in March 2026 and with every commit from a single developer, and PyPI shows about 1,900 downloads in the last month, against tens of thousands of stars for ahujasid's server.
Our take: if MCP for Blender is a skilled scripter you have to trust, blenderwright is a well-equipped assistant that mostly works through the tools you can see. That makes it the better default for anyone who would rather not let a model run Python in their scene.
Where it earns its place
An indie game developer making props for a Godot project
Block out and texture a set of low-poly props, then export them
Ask for each prop by description, run the mesh-quality check for non-manifold edges and loose vertices, unwrap UVs and export glTF. The structured tools mean no raw Python needs to be enabled; the cost is model tokens.
A product designer at a 10-person hardware start-up
Produce turntable renders of a new enclosure for a pitch deck
Import the STL, apply materials, use the studio lighting and turntable prompts and render with Cycles. Viewport screenshots let the model check framing before committing to a full render.
A hobbyist running a local model through Ollama
Automate repetitive scene clean-up on downloaded meshes
Set BLENDERWRIGHT_TOOLSETS=core or auto to keep schemas inside a 32,000-token window, then use the imported-mesh clean-up prompt to repair normals, decimate and fix scale. Free apart from local compute.
Before you commit
- Setup
- Install Blender 5.2 LTS, add the extension zip from GitHub Releases via Install from Disk, start the server in the N-panel, then add uvx blenderwright to the client (claude mcp add blenderwright -- uvx blenderwright for Claude Code). About 10 minutes.
- Integrations
- 276 tools, 13 prompts and resources for the scene, objects and materials over stdio. Works with Claude Code, Claude Desktop, Cursor, Codex and Ollama, which has a bundled chat client. Optional use of the Bool Tool, LoopTools and Node Wrangler extensions.
- Data and privacy
- Runs entirely on your machine with no telemetry; the README says it makes no network requests beyond the local socket to Blender. Scene data and screenshots go to your model provider. The add-on listens on 127.0.0.1:9876 with a token-checked handshake while the server is started.
- Cost at scale
- Free under MIT. The real cost is tokens: about 48,000 per request with every tool loaded, less with the core or auto toolsets.
- Limits
- Blender 5.2 LTS only; Blender must be open with the server started; one Blender per MCP server; viewport capture needs a visible viewport; the restricted Python namespace is not a security sandbox.
What’s good
- 276 validated tools across 32 modules
- Token-authenticated localhost socket
- Raw Python off by default, per session
- Toolsets cut context for local models
What’s not
- Blender 5.2 LTS only
- About 48,000 tokens of schemas with every tool loaded
- No sculpt strokes or grouped undo
- Young project with a single developer
Best for
3D artists and technical artists on Blender 5.2 who want an agent to model, light, animate and render through inspectable tools rather than free-form Python.
blenderwright alternatives
blenderwright: common questions
- Is blenderwright worth it?
- A deep, structured toolkit for driving Blender from a model, safer by design than script-running rivals, as long as you are on Blender 5.2 and can spare the context. We rate it 4.0 out of 5.
- How much does blenderwright cost?
- Free and open source (MIT). No paid tier, no telemetry and no external services.
- What are the downsides of blenderwright?
- Blender 5.2 LTS only; about 48,000 tokens of schemas with every tool loaded; no sculpt strokes or grouped undo; young project with a single developer.
- Who is blenderwright best for?
- 3D artists and technical artists on Blender 5.2 who want an agent to model, light, animate and render through inspectable tools rather than free-form Python.
- How much does blenderwright cost for a team or in production?
- Free under MIT. The real cost is tokens: about 48,000 per request with every tool loaded, less with the core or auto toolsets.
- Is blenderwright safe for company data?
- Runs entirely on your machine with no telemetry; the README says it makes no network requests beyond the local socket to Blender. Scene data and screenshots go to your model provider. The add-on listens on 127.0.0.1:9876 with a token-checked handshake while the server is started.
- What are the best blenderwright alternatives?
- The closest alternatives we have reviewed are MCP for Blender and OpenSCAD MCP.
Answers checked Nov 13, 2026.
Make blenderwright? Show this rating on your site
