ifc-console 0.1.4

A terminal interface that connects IFC files to LLMs. ifc-console loads your model with IfcOpenShell and serves it over MCP. Claude, Cursor, VS Code, Codex, and other MCP clients can inspect or edit the model while you keep control from the terminal.

No Blender or host BIM application is required. It runs locally on Windows, macOS, and Linux.

Terminal console	AI assistant	3D viewer

Why ifc-console

LLMs are good at BIM data work: finding elements, auditing properties, writing IfcOpenShell scripts, and explaining schema quirks. ifc-console is the safe, zero-setup bridge between an LLM client and an IFC file.

• You stay in control. The ask/edit switch decides whether the LLM may change the in-memory model. AI saving is disabled by default, so only you persist reviewed changes with /save.
• It runs anywhere. The core Python package runs on Windows, macOS, and Linux. The local 3D viewer is an optional install.
• It is honest. Errors include useful hints. Mutations are audited, saves are atomic with backups, and the docs state what the sandbox does and does not guarantee.

Install and run

uv tool install ifc-console
# Include the optional 3D viewer and browser chat:
uv tool install "ifc-console[viewer]"

You can use pip instead, or try the core application once with uvx ifc-console.

Start it in the folder containing your models:

cd path/to/your/models
ifc-console

Then:

> /file             choose an IFC model
> /connect codex    copy one-time client setup
> /viewer           open the optional 3D viewer

Paste the copied setup into your AI client and restart that client once. Future sessions are simply: start ifc-console, choose a file, and chat.

Ask or edit

Use /mode edit only when you want changes. Review them, then /save to keep them or /reload to discard them. The AI cannot change the mode, and AI tools cannot save unless you explicitly enable files.allow_ai_save.

On CPython 3.12+, eligible read-only generated code normally runs in a restricted process without network or subprocess access. Python 3.10 and 3.11 remain supported, but cannot provide that complete audit-hook boundary; auto reports a guarded fallback and strict refuses it. Mutating code runs in the main process because it must reach the live model. Read Safety before editing untrusted files or prompts.

What it includes

• IFC queries, properties, quantities, validation, clash detection, and CSV export.
• Multi-model coordination and an offline IFC/IfcOpenShell reference.
• A Python SDK for scripts, CI, workflows, and agent applications.
• Optional browser viewer, chat panel, and trusted operation plugins.

from ifc_console import Workbench

with Workbench.open("tower.ifc") as wb:
    walls = wb.query("IfcWall")
    print(len(walls))

Documentation

• Getting started
• Console and client setup
• Safety model
• Python SDK
• MCP tools and CLI
• Troubleshooting

For development setup and tests, see the contributing guide.

License

The core package is Apache-2.0 and uses IfcOpenShell (LGPL-3.0-or-later). The optional viewer includes Three.js (MIT) and web-ifc (MPL-2.0).

Inspired by Bonsai MCP.