AI-CAD

AI-CAD is a cleaned-up rebuild of an earlier CAD automation prototype. The first stable core focuses on offline DXF parsing, entity extraction, geometry normalization, and stone blank size analysis.

What Is Included

• DXF parsing with ezdxf
• Unified entity dictionaries for lines, arcs, circles, polylines, text, and inserts
• Layer/type filtering and drawing summaries
• Minimum oriented bounding rectangle analysis
• Multi-part stone grouping by endpoint connectivity
• JSON, CSV, and optional Excel export
• Automation-loop contracts for future computer-use and AutoLISP adapters
• AutoCAD current-viewport assistant contracts for MCP/AutoLISP integration
• A small CLI and focused pytest coverage
• An MCP server (aicad.mcp_server, built on FastMCP) exposing the DXF tools and, on Windows, timeout- and IsQuiescent-guarded AutoCAD COM tools plus a registry-driven catalog for driving external AutoLISP scripts

Heavy visual rendering and experimental tooling are intentionally left outside this rebuilt core. AutoCAD COM execution is now available as an optional, adapter-style layer through the MCP server (Windows + pywin32), kept behind automation contracts so the geometry/DXF core stays backend-independent. The current-viewport assistant direction is documented in docs/VIEWPORT_ASSISTANT.md, the general edit/validation loop in docs/AUTOMATION_LOOP.md, and migration notes in docs/MIGRATION.md.

Install

python -m pip install -e .

For Excel export:

python -m pip install -e ".[excel]"

For development:

python -m pip install -e ".[dev,excel]"

CLI

aicad summary path/to/file.dxf
aicad extract path/to/file.dxf --layer "stone-*"
aicad blank path/to/file.dxf --margin 2
aicad export path/to/file.dxf --format json --output output/entities.json

You can also run the package module directly:

python -m aicad summary path/to/file.dxf

Python API

from aicad import calculate_blank_size, extract_entities

entities = extract_entities("drawing.dxf", layer_filter="stone-*")
blank = calculate_blank_size("drawing.dxf", margin=2.0)

Every public API function returns a response dictionary:

{
    "success": True,
    "data": ...,
    "error": None,
    "metadata": {...},
}

MCP Server

aicad ships an MCP server (FastMCP) so agents can call the toolkit directly. Install the extra and run it:

python -m pip install -e ".[mcp]"
python -m aicad.mcp_server

Register it with an MCP client (e.g. a project .mcp.json):

{
  "mcpServers": {
    "ai-cad": { "command": "python", "args": ["-m", "aicad.mcp_server"] }
  }
}

Tools:

• DXF (cross-platform): aicad_extract_entities, aicad_blank_size, aicad_validate_dxf, aicad_build_viewport_probe, aicad_ping
• AutoCAD COM (Windows, optional): autocad_get_active_document, autocad_run_lisp, autocad_send_command, autocad_probe_viewport, autocad_diagnose_access — every COM call runs on a worker thread with a timeout and an IsQuiescent guard, so a busy AutoCAD returns a clean error instead of hanging the server.
• Script catalog: aicad_list_scripts, aicad_describe_script, autocad_run_script — a registry-driven layer for discovering and running your own AutoLISP commands through thin headless wrappers. The registry is read at call time, so adding a script needs no server restart.

Windows integrity note: the COM tools require AutoCAD and the MCP process to run at the same integrity level (both normal, or both "as administrator"). A mismatch (e.g. AutoCAD elevated, MCP not) blocks COM/ROT access and surfaces as an "invalid class string" error — autocad_diagnose_access reports this directly.

Repository Layout

src/aicad/          Core Python package
tests/              Reproducible tests and generated fixtures
docs/               Migration notes and project context
.codex/             Agent-facing context notes

Automation Direction

The intended editing loop has two modes:

• visible computer-use validation for screenshots, UI inspection, and interactive diagnosis
• quiet AutoLISP/headless automation for deterministic background edits and validation

Both modes should report through aicad.automation contracts so the geometry and DXF core stays independent from the CAD execution backend.

The next practical focus is the AutoCAD current-viewport assistant: Codex acts as the reasoning brain, while an MCP/AutoLISP backend reads the active view, selected entities, or a requested world-coordinate window and returns formatted JSON.