Blender MCP overview

Blender MCP 0.30.0 gives MCP clients typed, bounded control of a Blender session. It has two parts:

• The MCP server: a stdio server that publishes 14 tools and forwards

requests to the bridge.

• The Blender bridge: a Python module loaded inside Blender that executes

operations on Blender's main thread.

Everything runs on the local machine. The bridge binds only to IPv4 loopback and requires a random session token. There is no arbitrary Python tool, telemetry, external asset download or remote inference.

What it covers

• Inspection: scene and object metadata, mesh budgets, texel density,

directional UV stretch, material bindings, bone hierarchies, skin weights and linked clips.

• Object editing: primitives, custom meshes, transforms, origins, snapping,

names, sockets, duplicates, parents and collections.

• Mesh preparation: modifiers, UV unwrapping, decimation, LOD chains,

QuadriFlow remeshing, lightmap packing, texel-density correction, collision hulls, cleanup, booleans, joins, splits, mirrors, arrays and shells.

• Materials: PBR slot assignment with packed workspace textures, UV

transforms, alpha modes and material QA.

• Baking: normal and ambient-occlusion maps into a new workspace PNG.
• Assets: bounded GLB, FBX, OBJ and USD import; static, skinned and animated

FBX, GLB and USD export; Unity, Unreal and Godot axis presets; LOD groups, collision naming and bake pairs.

• Rigging: armature creation, automatic skin binding, weight repair, pose

authoring, action creation, animation QA and retargeting.

• Previews: isolated PNG previews, lit Eevee previews and multi-view review

sheets.

• Levels: bounded terrain meshes, bulk placement and surface snapping.
• Scatter: fixed Geometry Nodes instancing with explicit budgets.
• Ownership: edit leases shared by independent MCP clients.
• Evaluation: deterministic protocol and native application checks through

@cavi-ai/mcp-eval.

Tools lists every tool and action. Start with Inspect a scene or Import and export assets.

Evaluation

npm run eval:protocol checks discovery size, catalog token budgets, pagination, surface listing and protocol negotiation. npm run eval:core -- --blender /path/to/blender adds schema, structured output, repeated reads, contention and native application cases.

Native evaluation starts its own factory scene with a private connection and disposable workspace. It does not connect to an existing session. Reports stay under .tmp/evals/ and are ignored by Git. The evaluator does not invoke a model; results do not establish model task quality or game integration.

Boundaries

• Blender's operators and the render engine remain the source of truth. The

server reports refusals instead of simulating unsupported edits.

• Materials, shading and UV edits refuse shared mesh data. Make the mesh

single-user in Blender first.

• Exported assets still need the destination game's materials, scale, collision

and import settings.

• The optional image-to-3D worker produces a mesh; it does not rig, animate or

validate game import quality.

• Native checks target Blender 5.2.3 LTS. Behavior may differ on older releases.

The limits reference lists every input and output budget. The security model describes the transport, filesystem and ownership boundaries.

These docs were generated for release v0.30.0 from commit 9f9acb954ef6e12f98c51142950a8d9b923f9b5f.