MCP server for controlling Blender — create low-poly 3D models with AI
BlendBridge provides 8 tools with descriptions and input schemas, but exhibits significant gaps in parameter documentation, output schema definition, and error handling guidance. Most tool descriptions are 150-350 characters (within acceptable range) and explain the primary action, but parameter descriptions are sparse and output schemas are either missing or underdocumented. The server lacks explicit error recovery guidance, permission gates, and audit trail logging. Tool naming follows the verb_noun pattern well (execute_script, screenshot, set_viewport, render, get_scene_info, export_model, list_scripts, clear_scene), but several tools lack detailed parameter constraints and validation rules.
Remove all objects and clean up orphan data in the Blender scene. Clears meshes, materials, textures, and images with no users. Use this to start fresh before creating a new model.
Execute a Python script in Blender using the bpy API. This is the primary tool for creating and modifying 3D content. Write complete bpy scripts to create meshes, set materials, transform objects, etc. The script runs in Blender's Python environment with bpy already available. Two modes: - Inline: pass code directly via `script` (quick one-offs). - File-based: pass a path via `script_path` (for iterating — write the file with your editor, then execute by path. Edit specific lines between runs instead of rewriting everything). If both are provided, `script_path` takes priority.
Export the current scene to glTF/GLB format for use in game engines. The file is saved relative to the configured export base directory. Subdirectories are allowed (e.g., "weapons/sword.glb"). All transforms are applied before export.
Get information about all objects in the current Blender scene. Returns a list of objects with their names, types, transforms, materials, and mesh stats (vertex/face counts). Use this to understand what's currently in the scene before making changes.
List saved bpy scripts in the project scripts directory. Scripts are saved .py files that can be re-executed to recreate scenes. Use this to discover existing scripts you can iterate on or learn from.
Output schemas undocumented or incomplete. Tools like render(), get_scene_info(), screenshot() return JSON but the response structure is not explicitly documented in tool descriptions. LLMs cannot reliably extract fields (e.g., what fields does get_scene_info return? What does 'success' contain in execute_script output?).
Parameter descriptions lack detail and constraints. 'shading' param in screenshot() lists four valid enum values in the description but not as a formal enum constraint in the schema. 'rotation' in set_viewport() accepts 'list[float]' but does not specify length (must be 3?), range (degrees -180 to 180?), or units. 'timeout' in execute_script() lacks min/max bounds. Format and constraint information must be in schema, not just description text.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-22 | D | 53 | 2026-07-28+ | v2 |
| 2026-03-09 | F | 43 | - | v1 |
Render the current scene using Blender's render engine. Slower than screenshot but produces a properly lit, rendered image. Saves a PNG to disk. Read the file at the returned path to see the image.
Capture a screenshot of the active 3D viewport in Blender. Use this after execute_script to see what was created or modified. Saves a PNG to disk. Read the file at the returned path to see the image. Uses OpenGL viewport render — works reliably regardless of window focus.
Set the 3D viewport camera angle and position. Use this before screenshot to control what angle the model is viewed from. Supports preset angles or fully custom rotation.
No error recovery guidance. Tools return JSON with 'success: false' and 'error' fields, but do not guide the LLM on next steps. Example: execute_script() returns 'script_path must be an absolute path', but does not suggest how to construct an absolute path or what to do next. Error responses should include actionable recovery steps (e.g., 'Use os.path.abspath() or pass a full path starting with /').
Destructive/idempotent operations lack confirmation or dry-run support. clear_scene() is explicitly marked DESTRUCTIVE and permanently removes all objects and materials, yet offers no dry-run, confirmation step, or undo mechanism. The tool description does not warn of irreversibility or suggest when to call it. execute_script() with arbitrary Python code is inherently dangerous (code injection risk) and should enforce sandboxing or code review patterns.
Tool composition complexity. A multi-step workflow (create geometry with execute_script → adjust viewport with set_viewport → capture with screenshot) requires three sequential calls. While this is unavoidable for general-purpose bpy scripting, the server lacks a high-level 'create_and_render' or similar convenience tool that bundles common patterns and reduces round-trips.
No pagination or result limits documented. get_scene_info() returns 'a list of objects' with no mention of max count, pagination, or result size limits. If a scene has 1000+ objects, the response could be massive. No limit or offset parameters are offered.
Missing idempotency guidance. execute_script() does not document whether running the same script twice produces identical results or cumulative effects. If an LLM retries a script due to a transient error, will it create duplicate objects? Agents retry on ambiguous failures, non-idempotent tools risk side effects.
No audit trail or permission gates. Tools lack logging of who called what, with which parameters, at what time. No permission checks gate destructive operations (clear_scene, execute_script with arbitrary code). No scope declarations indicate required permissions. This violates compliance and debugging requirements.
Incomplete parameter descriptions for enums. render() accepts 'format' with values "GLB" or "GLTF_SEPARATE" but the description does not explain the difference or recommend one. set_viewport() lists presets (FRONT, BACK, LEFT, RIGHT, TOP, BOTTOM, THREE_QUARTER) but does not explain what each means visually or when to use them.