Skip to main content

Blender Bridge

Connect Profy to the Blender instance running on your own machine, so an expert can inspect the scene graph, execute Python (bpy), capture viewport screenshots to verify its own work, and export GLB. This is orthogonal to Profy’s cloud 3D capabilities. profy-visualize and profy-game-studio build web-runnable 3D with Three.js inside the sandbox. Blender Bridge operates on the actual .blend file sitting on your desktop.
Current implementation status (verified 2026-08-11): this plugin’s mcpServers declaration (uvx blender-mcp) has no consumer anywhere in the codebase. The PluginManifest.mcp_servers field in agent-runtime is parsed but never read, and Core’s MCP assembly path (buildCommunityMcpToolsAndConfig) only handles community plugin installations — it does not cover built-in manifests.In practice: ticking this plugin gives you its skill and prompt layer (the workflow discipline, safety rules, and export parameters below do take effect and shape model behaviour), but it does not give you callable Blender tools. Treat it as “a methodology for working with Blender” rather than “a remote control”, until the wiring is completed.

Declaration and activation

Two independent gates must both pass before the manifest contributes anything: The evaluation order is: selectable gate first, then every entry in requires (booleans compared by truthiness, lists by membership, strings by equality), then legacy conditions. Any failure short-circuits. The second gate means this plugin exists only inside Profy Desktop. A browser conversation in the cloud cannot select it and cannot activate it — the Blender you want to drive lives on your machine, and a cloud sandbox has no route to it.

Prerequisites

marketplace.installMode is ON_INSTALL, meaning setup happens at install time rather than lazily on first call. marketplace.availability is AVAILABLE and the category is creative-tools.

Workflow discipline (skill layer, active today)

The bundled blender-workflow skill defines the cadence the model follows. The core idea is look after every step:

Vision in the loop

The skill encodes one non-negotiable rule: screenshot after every single modelling operation, then ask whether the shape is right, the position is right, and the materials applied. Fix problems before moving on.
Never assume an operation succeeded — always verify visually.
This discipline does not depend on the MCP wiring. It constrains how the model self-verifies in any 3D task, which is why the skill layer is worth having even while the tool layer is incomplete.

Pre-flight checklist

  1. Verify the MCP connection by calling any Blender tool, e.g. get_scene_info
  2. Confirm the Blender version (4.x+) and the addon status
  3. Identify the current scene state before making any changes
Step 3 exists because “create a cube” means something different in an empty scene versus a scene that already has forty objects and a specific naming convention.

Capability inventory (available once MCP is wired)

Web export parameters

Models destined for the web have specific parameters, not “just export it”: The full pipeline is: model in Blender → optimise (decimate, bake textures) → export GLB → load in Profy’s visualize or game-studio → deploy via Sites.

Safety rules

Four hard constraints from the skill that the model will honour:
  • Never overwrite without asking — use “Save As” for new versions
  • Never delete objects without confirmation
  • Never modify files outside the project directory
  • Always work on a copy, never the original
These are worth reading as a statement of intent. An agent with bpy execution has, in principle, arbitrary file access through Python; the skill narrows that voluntarily. It is discipline, not a sandbox — if you need a hard boundary, keep the original file outside the project directory entirely.

Boundaries and failure modes

  • No Desktop, no plugin. It does not appear in the plugin panel in cloud conversations. That is the activation gate working, not a bug.
  • Blender not running means pre-flight fails. The skill requires an initial scene-inspection call to confirm connectivity. On failure the model should stop and report, not carry on pretending to model.
  • The MCP declaration is currently unwired (see the warning at the top), so “the model says it should screenshot but has no screenshot tool” is expected behaviour right now, not the model being lazy.
  • Polygon budget is on you. Web export has no automatic decimation fallback. An over-budget model still exports fine; it just loads badly.
  • uvx downloads on first use. If your machine has no network route to PyPI, launching the server will fail regardless of Blender’s state.

Troubleshooting

Verify your setup

  1. Open a new conversation in Profy Desktop. Blender Bridge should appear in the plugin panel — if it does not, Desktop is not connected and nothing further will work.
  2. Tick it and ask “what objects are in the current Blender scene?”. If you get methodology advice instead of actual scene data, you have hit the unwired state described at the top of this page.
  3. Ask it to outline the steps for “export the selected object as a web-ready GLB”. A correct answer mentions Draco, the 2048 texture ceiling, and applying transforms — which confirms the skill layer really did load.

3D and visualization

Cloud-side visualize / game-studio / AI 3D generation

Godot Bridge

Sibling bridge: game engine plus deterministic playtesting
Verified 2026-08-11. Sources: services/agent-runtime/src/plugins/builtin/blender-bridge/plugin.json, skills/blender-workflow/SKILL.md, prompts/BLENDER.md, services/agent-runtime/src/plugins/registry.py (_check_activation), services/agent-runtime/src/plugins/utils/types.py.