Summary

Assembling a scene stops being twenty tab switches.

Search, preview, download and import all happen in one place — Poly Haven for HDRIs and textures, Sketchfab for models, Rodin or Hunyuan3D when nothing off the shelf fits. Generation is asynchronous by design: generate hands back a job, a poll tool says when it is done, and a separate call puts the result in the scene.

What it is

A bridge into a running Blender over a JSON-over-TCP socket: the scene and its objects inspected, the viewport captured, Python executed inside Blender, and asset pipelines out to Poly Haven, Sketchfab, Hyper3D Rodin and Hunyuan3D. It comes in two halves — a Python MCP server you run, and an addon you enable inside Blender.

What you get

The current scene and any named object described back, and a screenshot of the viewport at a size you set
Shapes created, deleted and modified, and materials applied or created on an object
Arbitrary Python executed inside Blender — the escape hatch for anything the other tools do not cover (`execute_blender_code`)
Poly Haven browsed by category and searched by asset type — HDRIs, textures, models — then downloaded into the scene at a resolution and format you choose, with a downloaded texture applied to a named object
Sketchfab searched, a model's thumbnail previewed before you commit, then downloaded and scaled so its largest dimension matches a target
3D assets generated from text or from images by Hyper3D Rodin or Hunyuan3D, polled to completion, then imported under a name you give
A status check for each integration, so you know which ones are switched on before planning around them

Requirements

Uvx on your PATH and Blender open — every call reads or writes the scene that is loaded. Install the addon with `uvx blender-mcp install-addon`, enable **Interface: Blender MCP** in Blender's preferences, then open the BlenderMCP tab in the 3D viewport sidebar and click Start MCP Server. The server reaches Blender at `BLENDER_HOST` and `BLENDER_PORT`, which default to localhost and 9876. Run only one instance of the MCP server, not one per client. Asset integrations are toggled by checkbox inside Blender; their keys can be stored in the addon preferences or supplied as environment variables such as `BLENDERMCP_SKETCHFAB_API_KEY` and `BLENDERMCP_HYPER3D_API_KEY`. Two things to know going in: `execute_blender_code` runs arbitrary Python in Blender, so save your work before using it; and telemetry is on by default, switched off in the addon preferences or with `DISABLE_TELEMETRY=true`.

Setup effort

One command — uvx blender-mcp