Bonsai MCP

A local Model Context Protocol server that connects any MCP client (Claude Desktop, Claude Code, Cursor, VS Code, OpenAI Codex) to a running Blender + Bonsai (BlenderBIM) session. Inspect the scene, query the loaded IFC project, capture the viewport, and run Python inside Blender.

Documentation site: show2instruct.github.io/bonsai-mcp

MCP client  --stdio-->  bonsai-mcp  --127.0.0.1:9878-->  Blender add-on (bpy + Bonsai + IfcOpenShell)

Part of IFC-CoPilot: A Tool-Based Framework for LLM-Driven IFC Building Design.

News

• [2026-08] v1.2 released: three new tools that sync Blender after IFC edits without a full reload (refresh_view, refresh_geometry, reload_project), plus IfcOpenShell selector queries in list_elements. All releases.
• [2026-07] Bonsai MCP is public, with a documentation site: a lighter, code-generation-first rewrite of our earlier ifc-bonsai-mcp server.
• [2025-11] MCP4IFC paper released on arXiv, the study that motivated this server.

Relation to ifc-bonsai-mcp

This repository supersedes ifc-bonsai-mcp, which we used for the experiments in the MCP4IFC paper and which stays online for reproducibility.

The original ships a large set of predefined tools. This one is built around code generation. Modern LLMs write IfcOpenShell code well, so a small set of guarded code tools covers more IFC tasks than any fixed tool list, and adapts to tasks we never anticipated. The code-generation workflow is reworked and simplified here, and the server is lightweight, quick to connect, and easy to use.

Use this repository unless you need the original's fixed tools or the exact paper setup. It is the one we keep updated.

Requirements

• Blender 3.6+ (4.x recommended).
• Bonsai (BlenderBIM), only for the IFC tools. Install it from bonsaibim.org.
• uv, one-time install below. Runs the server and handles Python 3.10+ for you.
• git, to clone this repo. The server runs from your local checkout; it is not published to PyPI.

Setup (one-time)

Do these once. uv installs everything into the repo's own .venv the first time you run it; nothing is installed globally, and nothing is fetched from PyPI.

1. Install uv:

# macOS / Linux
curl -LsSf https://astral.sh/uv/install.sh | sh
# Windows PowerShell
powershell -ExecutionPolicy ByPass -c "irm https://astral.sh/uv/install.ps1 | iex"

2. Clone this repo somewhere permanent, and note its full path; you point your MCP client at it below:

git clone https://github.com/Show2Instruct/bonsai-mcp.git

Quick start

1. Install the Blender add-on

1. In Blender: Edit > Preferences > Add-ons. On Blender 4.2+/5.x, open the ▾ menu (top-right) > Install from Disk...; on older versions use the Install... button.
2. Pick blender_addon/bonsai_bridge.py from the repo you cloned (that exact file, not scripts/package_addon.py), or install the bonsai_mcp_bridge-X.Y.Z.zip attached to the matching GitHub release the same way. Then tick Bonsai MCP Bridge to enable it.
3. Open the sidebar (press N) > Bonsai MCP tab > Start Bridge. It now listens on 127.0.0.1:9878.

2. Connect your client

Point your client at the folder you cloned. Everywhere below, replace /path/to/bonsai-mcp with that path. On Windows in JSON, either use forward slashes or double every backslash (C:\\Users\\you\\bonsai-mcp).

• Claude Code:

  claude mcp add bonsai-mcp -- uv run --directory /path/to/bonsai-mcp bonsai-mcp
  

• Claude Desktop, Cursor, VS Code, and other clients: add this to the client's MCP config, then restart the client:

  {
    "mcpServers": {
      "bonsai-mcp": {
        "command": "uv",
        "args": ["run", "--directory", "/path/to/bonsai-mcp", "bonsai-mcp"]
      }
    }
  }
  

  On Windows, if the client can't find uv, use its full path as the command (e.g. C:\\Users\\you\\.local\\bin\\uv.exe).

uv run builds the project into the repo's .venv on first launch (editable, so your local edits take effect on the next restart) and caches it after that.

3. Verify

With Blender running and the bridge started, run this from inside the cloned folder:

uv run bonsai-mcp doctor

doctor pings the bridge and reports Blender / IfcOpenShell status and the tools it exposes.

Per-client config paths, the Windows uv.exe path note, and optional BONSAI_MCP_* settings are in docs/clients.md. Example configs live in examples/; set the --directory path to your clone.

Tools

Fourteen tools, each tagged [QUERY] (read-only) or [EDIT] (mutates state).

| Category | Tool | Purpose | | -------- | ------------------------- | ----------------------------------------------------------- | | QUERY | get_scene_info | Scene summary plus an optional filtered object list (paged). | | QUERY | get_selected_objects | Per-object info for the current selection (capped). | | QUERY | list_elements | IFC-backed elements filtered by class (inheritance-aware), name, storey, or an IfcOpenShell selector query (properties, materials, attributes); paged. | | QUERY | get_psets | IFC property and quantity sets for one or many objects (paged batches). | | QUERY | get_viewport_screenshot | Capture the viewport: view or azimuth/elevation, direction-aware fit, per-storey floor plans, shading incl. color-by-class, plus structured viewport state with depth. | | QUERY | get_ifc_project_info | Schema, counts, materials, classifications. | | QUERY | get_spatial_structure | Site -> building -> storey -> space tree with element counts. | | QUERY | get_quantities | Quantity takeoff (areas, volumes, lengths) by class, optionally per storey. | | EDIT | execute_ifc_code | Run IfcOpenShell / Bonsai API code. bpy blocked. | | EDIT | execute_blender_code | Run arbitrary Python with full bpy access. | | EDIT | refresh_view | Sync the scene after data-only IFC edits (names, psets); milliseconds, no disk I/O. | | EDIT | refresh_geometry | Rebuild geometry/placement for specific elements; targeted, no disk I/O. | | EDIT | reload_project | Full scene rebuild from the in-memory model (slow, explicit escape hatch). | | EDIT | save_ifc_file | Write the model to disk: in place or save-as (guarded). Durability only. |

Every tool returns structured content alongside readable JSON text, and the server also exposes MCP resources (bonsai://project, bonsai://scene, per-element psets) and two workflow prompts (model-audit, visual-verify).

Full reference with inputs, outputs, and examples: docs/tools.md.

Example prompt, with an IFC project open in Bonsai:

"List every wall in the model with its fire rating."

Safety

The bridge binds to 127.0.0.1 only and has no authentication, and execute_blender_code runs arbitrary Python. Treat it like an open Python REPL on your machine and never expose it to a network. See docs/safety.md.

Documentation

Rendered site: show2instruct.github.io/bonsai-mcp

• Installation
• Client setup (Claude, Cursor, VS Code, OpenAI)
• Tools reference
• Example prompts
• CLI reference
• Safety
• Troubleshooting

The changelog can be found on the GitHub Releases page.

Contributing

See CONTRIBUTING.md. Short version, with Python 3.10+ and uv:

uv venv
uv pip install -e ".[dev]"
uv run python -m ruff check src tests blender_addon scripts
uv run python -m pytest

Add tests and docs for behavior changes. Report security issues privately through GitHub Security Advisories (see SECURITY.md).

License

MIT. See LICENSE.

Disclaimer

This is not an official Bonsai release. Bonsai MCP is an independent project, not affiliated with or endorsed by the Bonsai (BlenderBIM) project, bonsaibim.org, or the Blender Foundation.