λ lambdacad-mcp

Let AI draw in your CAD. An MCP server for any AutoLISP-capable CAD — no COM, no SDK, no plugins. Pure AutoLISP. The reference adapter drives BricsCAD on Linux: the first MCP server for a professional DWG CAD on Linux.

"Draw a DN80 flange: front view, hatched section, dimensions — then model it in 3D."

Real-time capture, unedited: front view with bolt circle, hatched section, radial/diameter dimensions, leader, title block — then the same flange modeled as a 3D solid with an AI-written parametric AutoLISP loop and one boolean subtraction. ~20 seconds, 12 tool calls.

Complete 2D drawing, drawn by Claude	3D solids with booleans	Clean render for AI vision

Why this exists

Every other MCP server for a professional DWG CAD (multiCAD-mcp, autocad-mcp, CAD-MCP, ...) talks to the CAD through Windows COM — none of them can run on Linux, and none of them does real 3D.

BricsCAD is the professional DWG CAD that does run natively on Linux. This server drives it through a mechanism that needs no API at all: an AutoLISP file bridge. The MCP server writes LISP expressions to a file, triggers a tiny MCP command inside the CAD, and reads the result back. That's the whole trick — and it unlocks everything the CAD can do, including 3D solid modeling.

Because the bridge is plain AutoLISP + standard DXF entmake + AutoCAD-compatible commands, it is portable by construction: AutoCAD, ARES Commander (also Linux-native!), ZWCAD, GstarCAD and the IntelliCAD family all speak the same language. Porting means swapping the autoload file name and the keystroke trigger — adapters welcome.

MCP client (Claude, Cursor, ...)            BricsCAD (Linux)
        │                                        │
        ▼                                        ▼
  server.py (FastMCP, 128 tools)          c:MCP AutoLISP command
        │        writes LISP  ──►  /tmp/mcp_bricscad.in
        │        triggers via xdotool ──► reads, evals line by line
        │        reads result ◄──  /tmp/mcp_bricscad.out

What it can do

• Full 2D drafting — lines/arcs/polylines/splines, line weights per layer (0.50 walls / 0.35 edges / 0.18 dimensions — measured on the plotted page, not just set), real HATCH (with islands), dimensions incl. radius/diameter/angular, leaders, linetypes (chain/hidden), text styles, trim/extend/fillet/chamfer/offset/mirror/arrays, blocks
• Semantic shapes — slot, bolt_circle, rounded_rect, polygon, center_mark: one call, the server does the math
• 3D solid modeling — box/cylinder/sphere/cone/wedge/torus, extrude/revolve/sweep, union/subtract/intersect, slice, fillet/chamfer on solid edges, 3D move/rotate around any axis — and flatshot_view: auto-generate correct 2D drawing views from the 3D model
• Batch & atomic — draw_batch executes N operations in one round-trip, wrapped in one undo group (one batch = one Ctrl+Z)
• Talk about what the user selected — get_selection reads the GUI pickfirst set; select_by/select_window highlight what the AI found before acting; modify tools accept "selection"
• Verification without screenshots — bounding boxes, key points, exact lengths/areas, intersection checks (collision detection), solid volume/centroid
• Design review — tools that answer an engineering question, not just run a command (new in 1.2) — boolean_verified cuts and measures the change against the volume you expected (a cut that lands short reads SHORT, not success); overhang_report reads the STL mesh and says whether a part prints without supports, scoring all six orientations; contact_report over an assembly (assembly_place/assembly_list/assembly_clear) says who touches whom, who interferes, and who is an orphan — a part touching nothing, because its job was superseded and nobody noticed; hole_report physically probes every hole and says which are actually open (a drawing lies about its holes until something checks); graph_build/graph_trace read the 2D geometry as a connectivity graph — separate networks, exact run lengths, and the open ends that are real breaks vs T-junctions, with the joining tolerance emerging from the drawing instead of hardcoded
• The drawing REMEMBERS (new in 1.3, beta) — memory_enable puts a journal beside the .dwg that the CAD writes by itself, catching the edits the human makes in the GUI while no tool is running; tag_entity gives entities semantic roles ("muro-nord", "asta-principale") keyed to handles, which survive save and reopen; world_report answers what is here, what did I call it, what changed since I last looked — and revalidates every claim against the live drawing before answering (an entity erased behind the AI's back reads DEAD (erased since tagging), reported, never hidden). The next prompt — or the next session — starts knowing the drawing instead of re-surveying it
• Instructions in the protocol (new in 1.2) — the server ships the rules that cannot be learned by trying (OK is not evidence; calibrate before trusting; booleans keep the first handle; one line is one complete expression; consult the drawing's memory first), plus a lambdacad://traps resource that costs nothing until read
• Eyes when needed — render_view: BricsCAD's internal PNG render (clean, no UI, ~30 KB) for vision models
• Sheets, at a real scale — create_sheet (A4…A0, margins zeroed so paper mm = sheet mm), add_viewport (ask for 1:50 and it is 1:50 — measured back off the entity, and parked on a non-plotting layer), fit_viewport (measures what is drawn and picks the largest STANDARD scale that fits — 1:37 is not a scale), title_block (border + ISO 7200 cartouche, sized from the sheet), export_sheet_pdf, which checks its own work by measuring the page of the PDF it just wrote
• Deliverables — save DWG, export DXF, PDF and STL (export_stl: 3D solids straight to the slicer, parked on the origin)
• Several drawings open — list_documents shows which one the tools are pointing at (plus unsaved changes and read-only), use_document picks another; every reply says which drawing answered, and warns when it changes underfoot
• Escape hatch — run_lisp evaluates arbitrary AutoLISP (multi-line, multi-expression): capable models write parametric code and love it

Full reference: docs/TOOLS.md · How it works: docs/ARCHITECTURE.md

Measured, not promised

Details and model-behavior notes (Sonnet writes parametric LISP, Haiku uses typed tools): docs/BENCHMARKS.md

Tests

BRICS_DISPLAY=:0 python tests/test_tools.py     # ~135 checks incl. review + memory groups, ~2-4 min
python tests/test_tools.py 3d edges export      # or just some groups
python tests/test_tools.py --list

An integration suite: it drives the CAD you already have running, then measures what came out — closed-form volumes, bounding boxes, the value stored inside a dimension, triangle counts read out of the STL. Never "no error".

That distinction is the whole point. The bridge answers OK whenever the CAD accepted the command, which is not the same as the CAD having done what was asked. Every bug found so far was of that shape: the classic CHAMFER silently switching to all edges; a layout that plots rotated 90°, so an A3 landscape sheet comes out portrait, the drawing falls off the paper and the export writes a perfectly valid, entirely empty page; and — the one that reached a release — running object snap dragging programmatic coordinates onto nearby geometry, so five blocks placed in a row each landed on the previous. A volume, a bounding box, a measured dimension and the page of a PDF are exact signatures: every one of those surfaced as arithmetic that did not add up.

The suite draws in a scratch region (x ≥ 50000), erases it afterwards, and fails if the entity count does not come back to where it started. It still works in the active drawing, so point it at a scratch one.

Usage telemetry (local only)

With 128 tools, the interesting question is which ones earn their place — and which fail often enough to signal a docstring or design problem. The server appends one JSON line per call to ~/.lambdacad/usage.jsonl:

{"t": 1786526823.4, "tool": "draw_batch", "ms": 198, "err": 0}