Architecture

File Structure

server.py                    FastMCP 3.0 server — all tool/resource/prompt registrations
├── config.py                Settings class (pydantic-settings, env vars)
├── security.py              validate_path() + sanitize_command() + sanitize_lisp()
└── backends/
    ├── base.py              AutoCADBackend ABC + data models
    ├── ezdxf_backend.py     Headless DXF ops via ezdxf library
    └── com_backend.py       Live AutoCAD control via pywin32 COM

Dual-Backend Strategy Pattern

AutoCADBackend (in base.py) is an abstract base class. Both backends implement the full interface. The server creates exactly one backend at startup via _make_backend():

Check AUTOCAD_MCP_BACKEND env var
If auto or com on Windows: try COM → fall back to ezdxf
If non-Windows or ezdxf: use ezdxf

Data Models

Defined in base.py:

EntityInfo: {handle, type, layer, color, linetype, visible, properties}
LayerInfo: {name, color, linetype, lineweight, is_on, is_frozen, is_locked, is_current}
BlockInfo: {name, origin, attribute_count, entity_count, is_xref, description}
DrawingInfo: {name, full_path, saved, entity_count, layer_count, block_count, extents_min, extents_max, units, version, backend}
CommandResult: {ok, payload, error}

Backend Differences

Feature	ezdxf	COM
Platform	Any	Windows only
AutoCAD needed	No	Yes (running)
View/zoom	No-op	Full control
Screenshots	matplotlib render	Win32 window capture
Commands/LISP	Not supported	Full support
Undo	DXF snapshot stack	Native AutoCAD undo
Threading	`asyncio.to_thread` via `_async()`	`ThreadPoolExecutor` (single thread, STA)
File formats	DXF only	DWG + DXF

FastMCP Server

The mcp object is configured with:

Lifespan: autocad_lifespan — initializes backend, stores in ctx.lifespan_context["backend"]
Middleware: ErrorHandling → AuditLog (custom timing) → Timing → Logging
_backend(ctx): Helper that retrieves backend from context, raises ToolError if unavailable

Adding a New Tool

Add abstract method to AutoCADBackend in backends/base.py
Implement in EzdxfBackend — wrap sync calls with _async(func)
Implement in ComBackend — wrap COM calls with self._com(func)
Register in server.py: @mcp.tool(...) calling _backend(ctx).method(...)
Use _dc(result) to convert dataclass returns to dicts
Write tests in tests/

2.5 KiB Raw Blame History