ifc-console 0.1.2

A terminal interface to connect IFC files to LLMs. ifc-console loads your model with IfcOpenShell and serves it over MCP, so any LLM client (Claude Code, Claude Desktop, Cursor, VS Code, Codex) can query and edit it, while you stay in control from your terminal. No Blender, no host app.

Terminal console	AI assistant	3D viewer
The console to open an IFC file and decide what the AI is allowed to do.	An AI assistant connected to ifc-console, answering questions about the open model.	The optional viewer in the browser, showing the model tree, the 3D view, and element properties.

Install and run

Install from PyPI with uv or pip:

uv tool install ifc-console    # puts the ifc-console command on your PATH, viewer included
# or: pip install ifc-console

Or try it without installing: uvx ifc-console. To update later: uv tool upgrade ifc-console.

Start it in the folder with your models:

cd path/to/your/models
ifc-console

The MCP server comes up right away. Then, in the console:

> /file        pick a model from this folder
> /connect all one-time HTTP setup for every supported LLM client
> /copy codex  copy one complete client config to the clipboard
> /mode edit   let the AI change the model (ask = query-only, the default)
> /viewer      3D view in your browser
> /help        everything else

Wire up each client once. The token is stable per machine and no IFC path is stored in the client config, so your daily loop is just ifc-console, /file, and chat.

The mode switch

One switch, owned by you; the LLM cannot change it. Anything finer-grained (per-tool prompts, allowlists) belongs to your AI client.

Switch with /mode. Every mutation path is gated, saves are atomic with timestamped backups, and each session writes an audit log.

Honest caveat: the guards stop accidents, not a determined adversary. Treat edit mode plus untrusted prompts like running a stranger's script.

What the LLM gets

11 core tools: project info, spatial tree, selector queries, element details, property sets, schema docs, file list/open/save, and a gated Python execute_ifc_code power tool.

3 more while the viewer runs: read your click-selection, highlight elements, and screenshot the canvas so it can check its own work.

Every response is one JSON envelope with an actionable hint on failure. Full reference: MCP tools.

The 3D viewer

Type /viewer. It runs entirely on localhost behind your session token: click an element and the LLM knows what "this wall" means; it highlights elements back and takes screenshots. Edits refresh the view live.

Install from source (for development)

If you would like to work on the code itself, clone the repo and install from your checkout (needs git and uv):

git clone https://github.com/nbharathik/ifc-console
cd ifc-console
uv tool install .

To update later: git pull, then uv tool install . --force. For working on the code itself, see Development.

Docs

• Getting started
• The console and connecting clients
• Safety model and 3D viewer
• Development

License

Apache-2.0. Bundles three.js (MIT) and web-ifc (MPL-2.0, unmodified); uses IfcOpenShell (LGPL-3.0-or-later) as a library.

Acknowledgments

Inspired by Bonsai MCP. If you would like to work with the Bonsai viewer in Blender instead of a standalone terminal, check that project out.