Blender MCP Integration

Read the shared execution guidance once per task for autonomous decisions, evidence cadence, and long-task continuity.

Use the narrowest MCP layer that improves the task.

Choose the layer

1. Prefer Blender's official Lab MCP for Blender 5.1+ live scene inspection, screenshots, documentation lookup, rendering, navigation, and Python execution.
2. Use Blender Agent Studio's MCP for deterministic batch inspection, standardized geometry evidence, and bounded authored-scene renders. blender_render_scene preserves cameras, lights, volumes and color management and returns the first image inline; inspectOnly discovers scene settings before rendering. The bundled Poly Haven search/download tools also acquire verified CC0 textures and HDRIs without another add-on or API key; see the rendering workflow for material setup and standalone CLI use.
3. Consider ahujasid/blender-mcp only when the request needs its additional remote-host, Poly Haven, Sketchfab, or external 3D-generation integrations and accepts the extra installation, network, credential, and telemetry surface.
4. Do not run multiple add-on socket servers on the same host/port.

Read references/mcp-options.md before installing or replacing a Blender add-on.

Optional scene analysis runtime

Use blender_compare_reference for a camera-matched reference/silhouette/overlay board. It works directly through Blender without Rust and returns the image inline. A reviewed white-on-black mask optionally enables projection metrics; photos alone produce visual comparison, not guessed similarity scores. Author the reference camera with bpy or the live MCP before comparing geometry.

blender_describe_scene and blender_quality_report use the bundled Rust SceneIR analyzer. Run bun run setup:runtime in the installed plugin root with a stable Rust toolchain, or set BAS_RUNTIME_EXECUTABLE to a compatible built binary. Rebuild after updates. The existing tools remain independent of Rust.

Start with a compact scene description, then query an exact objectId and its descendants. Bounds use evaluated world geometry. Semantic roles come only from authored bas_role properties. Supply quality constraints from the brief; ground checks require named objects and groundZ. AABB relations are candidates, not exact mesh intersections. Keep the returned limitations and required visual questions attached to the findings. No numerical aesthetic score is produced.

Preserve reproducibility

• Keep the durable model in a Python source file even when using live MCP execution.
• Save a new .blend before risky arbitrary-code operations.
• Use official MCP screenshots for interactive iteration.
• Use $blender-agent-studio:blender-asset-validation for final clean-process inspection and fixed evidence cameras.
• Record which MCP tools were used in benchmark runs.

Benchmark MCP value

Compare:

1. plugin skills with MCP tools unavailable;
2. the same skills with the candidate MCP enabled.

Keep task, model, effort, time budget, and evaluator identical. Count execution failures, recovery turns, time, and final asset quality. Retain an MCP dependency only when it improves outcomes or materially reduces reliable completion cost.