MCP4IFC: IFC-BASED BUILDING DESIGN USING LARGE LANGUAGE MODELS

Interactive walkthrough: step-by-step creation of an L-shaped house through sequential prompts using MCP4IFC framework.

Abstract

Bringing generative AI into the architecture, engineering and construction (AEC) field requires systems that can translate natural language instructions into actions on standardized data models. We present MCP4IFC, a comprehensive open-source framework that enables large language models (LLMs) to directly manipulate Industry Foundation Classes (IFC) data through the Model Context Protocol (MCP). The framework provides a set of BIM tools, including scene querying tools for information retrieval, predefined functions for creating and modifying common building elements, and a dynamic code-generation system that combines in-context learning with retrieval-augmented generation (RAG) to handle tasks beyond the predefined toolset. Experiments demonstrate that an LLM using our framework can successfully perform complex tasks, from building a simple house to querying and editing existing IFC data. Our framework is re-leased as open-source to encourage research in LLM-driven BIM design and provide a foundation for AI-assisted modeling workflows.

Contributions

• An Open MCP Framework for IFC
  We introduce a comprehensive framework that enables LLMs to create, edit, and query IFC models through standardized Model Context Protocol (MCP) tool calls. This open, protocol-based approach ensures interoperability, allowing any MCP-compatible AI client to connect to our BIM tools.
• Dynamic Tool Expansion via RAG
  The framework moves beyond a fixed toolset by combining in-context learning with a Retrieval-Augmented Generation (RAG) system. This allows the LLM to dynamically generate new Python code for IFC-specific tasks, giving it the flexibility to handle novel operations and custom logic at inference time.
• End-to-end BIM manipulation
  We show that existing multimodal LLMs can successfully perform IFC querying, editing, and generation tasks using only natural language and tool calls, without direct user interaction with BIM software.

Method

The MCP4IFC system architecture (Fig. 1) is built on two core components: a Python-based MCP server that manages communication, and a Blender add-on that executes tasks. The server exposes the tool catalog to the LLM, while the add-on uses IfcOpenShell for all low-level IFC data operations and Bonsai to synchronize geometry with the Blender scene. This separation of concerns allows the LLM to operate on complex BIM data through a clean, standardized interface.

The MCP4IFC system architecture, showing the data flow from a user's prompt to the LLM Client, through the MCP Server, and into the Blender Environment for execution on the IFC Model.

A Comprehensive Toolset

The framework provides the LLM with a rich set of tools, divided into key categories:

• Scene Query & Context Tools: To understand the model's current state, the LLM can use tools that return a concise, structured JSON summary of the scene, including element GUIDs, IFC types, properties, and bounding boxes. It can also capture screenshots for visual feedback.
• Predefined Creation Tools: For common tasks, the agent can use high-level, parametric tools to create standard building elements like walls, slabs, roofs, doors, and windows, which are then correctly inserted into the IFC model.

Dynamic Code Generation

For any task not covered by the predefined tools, the agent can fall back on dynamic code generation:

• IFC Code Generation: The LLM can write Python scripts using the IfcOpenShell API. A RAG system provides up-to-date documentation and code examples to ensure the generated code is valid. This code is then executed in a secure sandbox.
• Custom Mesh Generation: For complex geometry, the agent can generate Python code using libraries like `trimesh` to create procedural meshes, which are then converted into IFC elements (e.g., `IfcBuildingElementProxy`).

Demo & Results

A short demonstration video will be available soon. In the meantime, see some generative results from our experiments below.

IFC Generation from Single Text Prompts

The LLM agent can generate complex IFC geometry from a single text prompt by invoking the framework's tools. (See Figure 4 in the paper).

For the full set of experiments and reproducible IFC files from the paper, check out our Google Drive archive.

BibTeX

@misc{nithyanantham2025mcp4ifcifcbasedbuildingdesign,
      title={MCP4IFC: IFC-Based Building Design Using Large Language Models}, 
      author={Bharathi Kannan Nithyanantham and Tobias Sesterhenn and Ashwin Nedungadi and Sergio Peral Garijo and Janis Zenkner and Christian Bartelt and Stefan Lüdtke},
      year={2025},
      eprint={2511.05533},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2511.05533}, 
}

Contact

For questions about our project, please contact Bharathi Kannan Nithyanantham or Tobias Sesterhenn.