new-acad: AI Agent for AutoCAD/Civil 3D

What changed

The new-acad project is an AI bridge designed to embed an intelligent agent directly within AutoCAD and Civil 3D. Unlike approaches that add CAD tools to an AI, new-acad integrates an AI agent that understands Civil 3D engineering practices into the CAD software itself. The agent exposes only two primary tools to the LLM: a plugin interface and MCP (Model Context Protocol) tools, with over 390 CAD methods managed by an engineering judgment layer rather than being directly exposed to the model. The system comprises three main components: a C3D plugin (C#) running within AutoCAD/Civil 3D, a Node.js-based panel and relay that acts as a central hub for translating natural language commands into plugin calls and relaying results back to the panel, and an MCP server offering 270 tools accessible by any MCP-compatible client. The project supports two AI modes: a default llm mode with a built-in LLM agent (OpenAI compatible) that requires only a user's API key, and an openclaw mode for connecting to self-hosted agent platforms. The latest release, v1.7.0, introduces significant enhancements, including verifiable decision chains with structured logging and replay capabilities, improved handling of large drawings with context summarization and structured cropping, and specialized tools for earthwork calculations such as corridor earthwork computation, grading optimization, and site mass haul analysis. The earthwork computation for corridors, for instance, shows a difference of only 0.1% compared to Civil 3D's surface volume method. The system also features a new bitmap-to-CAD tracing capability (v1.6.0) that can convert images into vector lines based on color regions, arc fitting, or skeletonization, and an image/drawing triage system that intelligently extracts dimensions before tracing. The interface language can now be set during installation and switched within the panel, supporting Chinese and English.

Why it matters for builders

This project offers CAD professionals a novel way to interact with complex engineering software using natural language. Builders can directly command AutoCAD and Civil 3D to perform intricate tasks like generating surfaces, designing corridors, or managing pipe networks without needing to learn specific command-line syntax or navigate complex menus. The integration of an independent agent that does not rely on external cloud platforms, requiring only an API key, provides flexibility and control over data. The focus on engineering practice and the inclusion of specialized tools for earthwork and design optimization can significantly streamline workflows and reduce manual effort.

Practical impact

Users can install new-acad via a provided setup executable or a green version zip. After installation and configuring an LLM API key, users can open Civil 3D, and the plugin will automatically load. They can then issue commands directly in the panel, such as "Raise the selected polyline by 0.5 meters." The project provides comprehensive documentation, including an ONBOARDING.md file specifically for AI agents, detailing environment checks, tool invocation, and troubleshooting. For developers, the project structure is clearly laid out, with separate directories for the C# plugin, Node.js server components, and documentation. The ability to compile the plugin and package distribution files is also detailed, requiring Civil 3D's development assemblies. The MIT license allows for broad use and modification.

Caveats and source limits

The project is not an official Autodesk product and has no affiliation with Autodesk. Users need to provide their own licensed versions of AutoCAD/Civil 3D (2025/2026) and an LLM API key. Some advanced Civil 3D functionalities, such as editing grading groups or parcel lines, are not yet exposed via API and thus cannot be controlled by the AI. The plugin DLL is locked while Civil 3D is running, requiring the CAD software to be closed for updates or recompilation. The system relies on Server-Sent Events (SSE) for real-time panel responses, with a 10-second polling fallback for connection status. The source does not provide independent benchmark results for the AI's performance on complex tasks beyond the stated 0.1% difference in earthwork calculations. Pricing information for any potential commercial use or support is not detailed.