The design treats a screenshot as interpretation and the measure and assert tools as the truth layer, which is what keeps a modelling session from drifting on the model's own description of its work. Start every session with `router_set_goal` — the guided surface is deliberately small, and if a tool is not visible you are expected to find it with `search_tools` and run it through `call_tool` rather than guessing the name.
A Python MCP server plus a Blender addon. The server handles routing, validation and discovery; the addon runs the operation on Blender's main thread, which is where context-sensitive operators actually work.
- A small guided entry surface — `router_set_goal`, `router_get_status`, `browse_workflows`, `search_tools`, `call_tool` — so the model is not flooded with the whole catalog
- Task-sized macro tools such as `macro_cutout_recess` for openings, `macro_relative_layout` for placement, and `macro_finish_form` for bevel, subdivision and solidify passes
- Grouped scene tools under readable names: `check_scene`, `inspect_scene`, `configure_scene`
- Deterministic verification via the `scene_measure_` and `scene_assert_` families, so a result is checked rather than eyeballed
- Read-only spatial artifacts: `scene_scope_graph`, `scene_relation_graph`, and `scene_view_diagnostics` for projected extent, frame coverage and occlusion verdicts
- Reference-image workflows: `reference_images` to attach one, then staged comparison and correction against it
Python 3.11+ and the `blender-ai-mcp` package, plus the Blender addon that executes the operations. Vision is pluggable: a local MLX runtime by default, or an external provider selected with `VISION_EXTERNAL_PROVIDER` and, if needed, `VISION_EXTERNAL_CONTRACT_PROFILE`. Prompt-only clients can turn on the prompt bridge with `MCP_PROMPTS_AS_TOOLS_ENABLED`.
One command — docker run -i --rm \ -v /tmp:/tmp \ -e BLENDER_AI_TMP_INTERNAL_DIR=/tmp \ -e BLENDER_AI_TMP_EXTERNAL_DIR=/tmp \ -e ROUTER_ENABLED=true \ -e MCP_SURFACE_PROFILE=llm-guided \ -e BLENDER_RPC_HOST=host.docker.internal \ ghcr.io/patrykiti/blender-ai-mcp:latest