CadGPT

Status: Active development

CadGPT turns ChatGPT into a drawing-centric AutoCAD execution environment. ChatGPT remains the reasoning surface; CadGPT provides the local control plane, drawing binding, AutoLISP/Job orchestration, registry discovery, diagnostics, the AutoCAD add-in, and the CAD MCP bridge.

ChatGPT / CadGPT plugin
        ⇅
OpenAI Secure MCP Tunnel
        ⇅
CadGPT slim admission/control plane
        ↓
WorkRegistration + ToolLease
   ├── FILE
   └── CAD / HYBRID
          ↓
        CAD MCP
          ↓
       AutoCAD

The AutoCAD add-in can host ChatGPT in a dockable palette, but the add-in UI is not CAD authority. CadGPT runtime/work binding remains authoritative.

Current product contract

• explicit launch through the CadGPT/CG plugin or literal @cadgpt / @cg;
• one active work binds one drawing;
• current CAD facts are read fresh from the bound drawing;
• while CG is active with a bound drawing, requests such as draw / redraw / create / add / modify / vẽ / vẽ lại / tạo / sửa target AutoCAD by default;
• external image generation is used only when the user explicitly asks for an image/render/illustration rather than CAD output;
• AutoCAD is never launched implicitly by the MCP runtime;
• CAD MCP sleeps until actual CAD demand;
• managed user/runtime state lives under %LOCALAPPDATA%\CadGPT;
• file mutation uses canonical absolute paths inside approved roots;
• production runtime does not self-modify CAD MCP source;
• Revit development is not part of this repository. RevitGPT is maintained as a separate project/repository.

Launch and drawing binding

CadGPT reads the lightweight tray snapshot before starting full CAD MCP.

CG / @cg / @cadgpt
├─ 0 drawings
│  → Welcome screen
│  → WORK IDLE
│
├─ exactly 1 drawing
│  → auto-bind that drawing
│  → Workspace Ready
│
└─ 2+ drawings
   → Welcome screen + drawing list
   → user chooses exactly one
   → Workspace Ready

cg/list refreshes the tray-backed drawing list and does not wake full CAD MCP by itself.

A bound workspace follows:

1 work = 1 drawing

CadGPT never treats ambient AutoCAD ActiveDocument as drawing authority.

Commands

cg/       full command menu
cg/list   refresh open drawings
cg/cl     create or repair Lisp
cg/cj     create or repair Job
cg/job    list registered Jobs
cg/kug    update global HVAC knowledge (KUG)
cg/kud    update bound-drawing HVAC knowledge (KUD)
cg/mcp    CAD MCP development workflow (development builds only)
cg/help   help
cg/stop   stop current work

@cadgpt / @cg activate CadGPT; they are not the CLI namespace.

HVAC domain Knowledge Updater (KUG / KUD)

The Knowledge Updater stores HVAC workflow/domain knowledge separately from CadGPT internal CAD API, AutoLISP and MCP development knowledge.

• KUG / cg/kug — global reusable HVAC practice in %LOCALAPPDATA%\\CadGPT\\knowledge\\.md. Works without AutoCAD.
• KUD / cg/kud — bound-drawing HVAC knowledge in %LOCALAPPDATA%\\CadGPT\\drawings\\\\knowledge\\.md. Requires a valid, explicitly bound drawing workspace.

The dedicated knowledge_list, knowledge_read, and knowledge_upsert tools handle scope, revision checks (SHA-256), atomic writes and readback. For drawing knowledge, the exact binding/anchor is revalidated before access. The model should compare existing content, show changes and request human approval before updating. Unresolved HVAC facts stay unresolved.

CAD API/Lisp authoring, the Lisp Writer, and MCP Fixer remain in their respective internal knowledge and skills. KUG/KUD must not edit them. See skills/knowledge-updater/SKILL.md.

AutoCAD add-in

The managed add-in lives at:

addins/cadgpt-autocad/

It embeds ChatGPT in an AutoCAD dockable palette while leaving admission, work authority, and drawing binding in the CadGPT runtime.

Install or refresh it after closing AutoCAD:

cadaddin.bat

Then open AutoCAD and run:

CADGPT

The installed bundle is placed under the current user’s Autodesk ApplicationPlugins directory and is configured for AutoCAD startup loading. The panel itself stays hidden until requested.

See addins/cadgpt-autocad/README.md for host probing, build details, WebView2 profile behavior, and compatibility notes.

CAD intent and live state

While CadGPT is active with a drawing workspace:

• drawing/editing language targets the bound AutoCAD drawing by default;
• an attached image may be used as a geometric/visual reference for CAD work;
• the attachment does not change the output destination to image generation;
• layer counts, entity counts, geometry, properties, selections, and other current-state questions require a fresh CAD read in the same turn;
• conversation history is never treated as live CAD state.

Curated CAD operating knowledge lives in:

knowledge/cad/WORKING_KNOWLEDGE.md

Diagnostics

Real product failures and incorrect behavior are accumulated separately from model knowledge:

diagnostics/cad/ERROR_LOG.md

Use the error log for tool/runtime failures, routing errors, latency regressions, workarounds, unexpected CAD behavior, and other defects discovered during real use.

Only stable CAD operating knowledge should be promoted into knowledge/cad/WORKING_KNOWLEDGE.md.

Registry

CadGPT exposes one effective registry with strict ownership.

Internal Registry
├─ MCP tools
├─ system skills
├─ bundled Lisp
└─ official Internal Jobs

User Registry
├─ Lisp capabilities
└─ user Jobs

User Registry cannot overwrite Internal Registry.

TBH Toolkit

Bundled TBH content lives under:

resources/cad/internal-lisp/tbh-toolkit

The official Internal Direct Job tbh loads:

resources/cad/internal-lisp/tbh-toolkit/tbhloader.lsp

Per drawing:

tbhloader.lsp load success → TBH = ON
tbhloader.lsp load failure → TBH = OFF

A failed Internal Direct Job is surfaced as a failure instead of being bypassed through lower-level tools.

AutoLISP

User-managed Lisp lives under:

%LOCALAPPDATA%\CadGPT\libraries\lisp

Editing uses controlled checkout → validate → real CAD test → promote workflows. Generic file tools cannot mutate permanent managed libraries.

System Lisp authoring guidance lives under:

skills/write-lisp/

Jobs

User-created/imported Jobs live in managed AppData and User Registry.

CadGPT supports:

• Direct Jobs — fixed local implementations without model planning;
• Reasoning Jobs — sequential READ → PLAN → REVIEW → EXEC → READBACK workflows.

When a reusable Job needs a dynamic form of an already-working Lisp, CadGPT seeds a byte-for-byte copy once under the Job’s own dynamic-lisp/ folder, patches only declared data/sections with hash-guarded exact replacements, and reuses that persisted copy on later runs. Shared Lisp logic stays in the Lisp Library; Job-owned derivatives are not adapter commands.

Job authoring guidance lives under:

skills/jobcreate/
knowledge/jobs/

Managed AppData

Production data lives under:

%LOCALAPPDATA%\CadGPT\
├── libraries/
│   ├── lisp/
│   └── jobs/
├── registry/
│   └── user/
├── workspace/
│   ├── lisp-draft/
│   └── job-draft/
├── runtime/
├── data/
├── drawings/
│   └── /knowledge/
├── knowledge/                 # Global HVAC domain knowledge
├── state/
└── logs/

CADGPT_APPDATA_ROOT is a development/test override only and must not point to a repository-root appdata folder.

Installation and daily use

Requirements:

• Windows
• Node.js 20+
• Python 3.11–3.14
• AutoCAD for live CAD work

One-time source setup:

setup.bat

Common runtime commands:

run.bat status
run.bat restart
doctor.bat
acceptance.bat

Install/refresh the AutoCAD add-in:

cadaddin.bat

After setup, the Windows tray hosts the lightweight CadGPT control plane and Secure MCP tunnel.

Runtime safety

Important invariants:

• explicit CadGPT session launch;
• execution authority comes only from the current work_handle;
• FILE and CAD authority remain explicit;
• one work binds one drawing;
• stale/reopened drawing lifetimes do not silently reuse old drawing authority;
• destructive CAD actions use guarded preview/execution contracts;
• work authority expires after idle timeout;
• different ChatGPT conversations do not intentionally share work authority.

Development-only CAD MCP improvement

Source/development builds may expose cad-mcp-dev.

Its writable scope is limited to:

runtimes/cad-mcp/**

It is not part of the production self-modification surface.

Repository layout

src/cadgpt/                    CadGPT orchestration/control plane
runtimes/cad-mcp/              active AutoCAD MCP runtime
addins/cadgpt-autocad/         AutoCAD dockable ChatGPT panel
resources/cad/                 bundled CAD resources / TBH Toolkit
skills/                        system authoring/development skills
knowledge/                     curated runtime/model knowledge
diagnostics/                   product error evidence
registry/                      registry contracts/docs
scripts/                       setup/build/audit/install utilities
tests/                         regression suite

Historical Stage 0 experiment plans and the old Revit MCP preservation area are intentionally not kept in the active CadGPT tree. Git history remains the archive for superseded implementation material.