debim

🎯 Why debim?

"debim originated from a simple desire: to make AI calculate accurate Bills of Quantities (BOQ). Asking an LLM to guess building dimensions directly in text leads to fatal hallucinations. An exact mathematical model (BIM) is essential, yet traditional IFC files are bloated and overwhelm AI context windows. The solution is Declarative YAML — but the resulting Building-as-Code engine proved far more transformative than our initial goal."

Traditional BIM tools (like Revit or Archicad) were conceived over 25 years ago for humans clicking with computer mice. They lock architectural data inside heavy, proprietary gigabyte files (.rvt), charge thousands of dollars in annual licenses, and remain completely opaque to modern automation and AI agents.

debim is built on 6 Core Architectural Tenets:

1. Building-as-Code & Git-Native: Buildings are software. Expressed as compact YAML/.debim (Kilobytes, not Gigabytes) for transparent Version Control, line-by-line Git diffs, and branching.

2. Deterministic Code Compliance: Building codes and engineering regulations are treated as automated Unit Tests (pytest), catching setback violations and structural errors in 0.01 seconds before ground is broken.

3. Zero-License & Zero-Friction Visualization: Instant geometric verification through lightweight 3D HTML viewers that load in any browser or mobile device in 2 seconds without expensive licenses.

4. Universal Bridge & Dual Representation: Seamlessly connects 2D drafts, 3D DCC tools (Blender/SketchUp), and open IFC standards using a dual approach: 90% geometric primitives for engineering/BOQ + 10% baked GLB assets for architectural refinement.

5. Human & AI Super-Collaboration: Designed with explicit uncertainty flags (review_status: needs_review), enabling humans and autonomous AI agents to co-author and verify building models without friction.

6. "The Mermaid.js of BIM" (Zero-Dependency & Air-Gapped Sandboxes): Operates seamlessly inside secure enterprise environments, air-gapped VMs, and LLM Code Interpreters (ChatGPT, Claude Artifacts, Gemini Code Execution) using 100% Python Standard Library with zero pip dependencies.

Related MCP server: ifc-mcp

🤖 Autonomous AI-Agent Setup

If you use an AI coding assistant (like Antigravity, Cursor, Claude Code, Jules, or ChatGPT/Copilot), you don't even need to install it manually!

Just copy and send this prompt to your AI:

"Please read https://github.com/PRIDA-TAKON/debim and AGENTS.md, install debim in my environment, and run debim --help to verify."

Your agent will inspect the repository, install the dependencies, and verify everything automatically.

🔌 Model Context Protocol (MCP) Server

debim natively implements the official Model Context Protocol (MCP) via FastMCP (Python SDK). It provides LLMs and AI Agents (such as Claude Desktop, Cursor, Cline, Windsurf, Devin, and Antigravity) with deterministic tools to model, inspect, calculate, compile, and visualize buildings directly via function calling.

Connecting to Claude Desktop / Cursor / Cline

Add debim to your MCP configuration (claude_desktop_config.json or .cursor/mcp.json):

{
  "mcpServers": {
    "debim": {
      "command": "debim",
      "args": ["mcp"]
    }
  }
}

Or run via Docker:

{
  "mcpServers": {
    "debim": {
      "command": "docker",
      "args": ["run", "-i", "--rm", "ghcr.io/prida-takon/debim:latest"]
    }
  }
}

Exposed MCP Tools

Running the MCP Server Locally

# Start MCP server over stdio
debim mcp

🚀 Quickstart

1. Installation

debim is officially available on PyPI and requires Python 3.11+.

Choose your operating system and preferred terminal:

🪟 Windows (PowerShell / Command Prompt)

# Option A: Isolated CLI (Recommended via pipx or uv tool)
pipx install "debim[all]"
# or: uv tool install "debim[all]"

# Option B: Standard pip
python -m pip install --upgrade pip
pip install "debim[all]"

# Option C: Windows Package Manager (Zero-Python / One-Click Native)
# Note: Manifest is currently pending upstream review on microsoft/winget-pkgs
# winget install debim

# Verify installation
debim --help

🍎 macOS (Terminal / zsh)

# Option A: Native Homebrew Tap (Recommended on macOS)
brew install PRIDA-TAKON/debim/debim

# Option B: Isolated CLI (via pipx)
brew install pipx
pipx ensurepath
pipx install "debim[all]"

# Option C: uv tool
uv tool install "debim[all]"

# Option D: Virtual Environment
python3 -m venv .venv
source .venv/bin/activate
pip install "debim[all]"

# Verify installation
debim --help

🐧 Linux (Ubuntu / Debian / Fedora / Arch)

# Option A: Isolated CLI (Recommended on Ubuntu 23.04+ / Debian 12+)
sudo apt update && sudo apt install -y python3-pip python3-venv pipx
pipx ensurepath
pipx install "debim[all]"

# Option B: Virtual Environment
python3 -m venv .venv
source .venv/bin/activate
pip install "debim[all]"

# Option C: uv tool
uv tool install "debim[all]"

# Verify installation
debim --help

📦 Package Extras Summary

🐳 Docker Container (Zero-Install & Sandbox)

If you prefer not to install Python locally or want to run debim in an isolated sandbox for AI agents:

# Pull official pre-built image from GitHub Container Registry (GHCR)
docker pull ghcr.io/prida-takon/debim:latest

# Run debim MCP Server over stdio for AI Agents / Claude Desktop
docker run -i --rm ghcr.io/prida-takon/debim:latest

# Or execute CLI commands in your current directory
docker run --rm -v "${PWD}:/workspace" -w /workspace ghcr.io/prida-takon/debim:latest debim --help

# Alternatively, build the container image locally from source
docker build -t debim .

🔒 Air-Gapped Sandboxes & Zero-Dependency Execution (Pure Standard Library)

In highly secured enterprise environments, air-gapped VMs, or LLM Code Interpreter sandboxes (e.g., ChatGPT Advanced Data Analysis, Claude Artifacts, Gemini Code Execution, Kaggle) where internet access and pip install are restricted:

# Run standalone zero-dependency engine (Pure Python Standard Library: math, json, re, argparse)
# Requires ZERO pip packages:
python skills/debim/scripts/debim_zero.py validate -m model.debim
python skills/debim/scripts/debim_zero.py qto -m model.debim

🛠️ Install from Source (Development)

git clone https://github.com/PRIDA-TAKON/debim.git
cd debim
pip install -e ".[all,dev]"

2. Basic Commands (v1.0.0)

# Initialize a new project
debim init my-project

# Validate schema syntax, structural continuity & bounding-box clash detection
debim validate -m examples/farnsworth_house/project.debim

# Run automated building code compliance checks (pytest)
debim test

# Calculate Quantitative Take-Off (Concrete m3, formwork m2, rebar kg)
debim qto -m examples/farnsworth_house/project.debim

# Generate official Thai Government BOQ (CGD ว480 with Factor F & steel density)
debim boq --rules cg-v480 --province nakhon_ratchasima -m project.debim -o dist/boq_v480.csv

# Generate project-scoped price template with international classifications
debim cost template -m examples/farnsworth_house/project.debim -o prices.template.yaml

# Estimate project budget & export commercial BOQ to CSV
debim cost -m examples/farnsworth_house/project.debim -p examples/farnsworth_house/prices.yaml -o dist/boq.csv

# Scaffold a new BIM element class boilerplate
debim scaffold element IfcRailing

# Preview 3D model in your browser (Three.js with Hierarchical Layer Explorer)
debim view -m examples/farnsworth_house/project.debim

# Compile declarative YAML to standard IFC4 building model
debim compile -m examples/farnsworth_house/project.debim -o dist/farnsworth_house.ifc

# Ingest Open GIS (Overpass BBox or GeoJSON) into native .debim urban model
debim urban ingest --bbox "13.734,100.558,13.742,100.566" -o dist/asoke.debim
debim urban ingest --geojson road.geojson --crs 32647 -o dist/road.debim

# Compile urban & civil infrastructure to ISO 19650 IFC4.3
debim compile -m dist/asoke.debim -o dist/asoke.ifc

# Launch Model Context Protocol (MCP) server over stdio
debim mcp

🏗️ Example project.yaml

schema: IFC4-Minimal
project:
  id: PRJ-2026-001
  name: "Townhouse-Feasibility"
  units: { length: METER, area: SQUARE_METER, volume: CUBIC_METER }

# 1. Spatial Structure
spatial_structure:
  storeys:
    - id: L1
      name: "Level 1"
      elevation: 0.00
      height: 3.50
    - id: L2
      name: "Level 2"
      elevation: 3.50
      height: 3.20

# 2. Reference Grid Axes
grids:
  axes_x: { A: 0.00, B: 4.00, C: 8.00 }
  axes_y: { 1: 0.00, 2: 5.00, 3: 10.00 }

# 3. Materials
materials:
  - id: CONC_240
    name: "Concrete 240 ksc"
    category: concrete
    unit_cost_ref: "MAT-CONC-01"
  - id: AAC_75
    name: "AAC Block 7.5cm"
    category: masonry
    unit_cost_ref: "MAT-AAC-01"

# 4. Elements
elements:
  # Column placed at grid intersection [A, 1]
  - class: IfcColumn
    tag: C-A1
    material: CONC_240
    profile: { shape: BOX, width: 0.20, depth: 0.20 }
    placement:
      grid: [A, 1]
      base_storey: L1
      top_storey: L2
    reinforcement:
      main: "4-DB16"
      stirrups: "RB6 @ 0.15m"

  # Beam spanning between [A, 1] and [B, 1]
  - class: IfcBeam
    tag: B-A1_B1
    material: CONC_240
    profile: { shape: BOX, width: 0.20, depth: 0.40 }
    placement:
      from_grid: [A, 1]
      to_grid: [B, 1]
      storey: L2
    reinforcement:
      main_top: "2-DB16"
      main_bottom: "3-DB20"
      stirrups: "RB9 @ 0.15m"

  # Wall with door host-child relationship
  - class: IfcWall
    tag: W-A1_A2
    material: AAC_75
    thickness: 0.075
    height: 3.10
    placement:
      from_grid: [A, 1]
      to_grid: [A, 2]
      storey: L1
    children:
      - class: IfcDoor
        tag: D1
        dimensions: { width: 0.90, height: 2.00 }
        offset_distance: 1.20

🔬 Empirical Research & Benchmark

debim prioritizes engineering precision and reproducible open science on our Kaggle Cloud Multi-Core Benchmark Suite:

1. ⚖️ 3D Visual Regression & Alignment Benchmark (255 Real-World Buildings)

Evaluating blind 3D geometric fidelity against ground-truth IFC models using Geometry Variant Deduplication + Balanced Macro-Averaging across 4,695 representative building elements:

• 🎯 85.50% Median Visual Fidelity: Surpassing the international standard benchmark ($\ge 85%$) across the majority of test suites.

• 🏗️ 89.07% Structural Match: Primary load-bearing elements (columns, beams, slabs, foundations) maintain Grade-A+ geometric alignment.

• ⚡ 82.64% MEP System Match: Ductwork, drainage, piping, and electrical fixtures align accurately in 3D coordinate planes.

📈 Progression Across Waves:

👉 Want to inspect raw visual data or reproduce tests yourself? Explore the full dataset and code on the Kaggle Benchmark Notebook.

2. 📦 Roundtrip Retention & Storage Reduction Study (407 Real-World OpenBIM Models)

Rigorously benchmarked against 407 real-world projects across architectural, structural, and complex hospital MEP domains:

• 🎯 100.0% Median Retention Rate: Extract and re-compile back to standard IFC4 without element loss.

• 📊 85.0% Mean Retention Rate: +26.1% retention gain across diverse real-world models (up from 58.9% in Round 1).

• 📦 88.4% Average Storage Reduction: Compresses raw IFC files by an average of 88.4% into compact declarative YAML.

• ⚡ 558M+ LLM Tokens Saved: Prevented 558,629,804 tokens from cluttering agent context windows.

• 🛡️ 100.0% Modern Schema Crash-Resilience: Zero fatal crashes or unhandled exceptions across standard IFC2X3, IFC4, and IFC4X3 datasets.

📈 Round 1 vs. Round 2 Stress Test Evolution:

📖 Read Full Research Paper: debim: An Empirical Study of Declarative Building-as-Code on 407 Heterogeneous Real-World OpenBIM Models
📦 Kaggle Public Benchmark Dataset: debim-5000-ifc-benchmark
⚡ Kaggle Automated Stress Test: debim-ifc-stress-test

❓ Frequently Asked Questions (FAQ)

📄 License & Attribution

• Software Code: Licensed under the MIT License — Copyright (c) 2026 Prida Takon.

• Research & Technical Reports: Licensed under Creative Commons Attribution 4.0 International (CC BY 4.0).

• Benchmark Datasets & Sample Models: Test fixtures in tests/fixtures/ originate from buildingSMART International and the Open IFC Model Repository under CC BY 4.0 / CC-BY-3.0.