agentcad

Product Introduction

1. Definition: AgentCAD is an open-source Model Context Protocol (MCP) server and Command-Line Interface (CLI) that functions as a specialized tool for AI coding agents. It falls into the categories of AI-assisted CAD software, programmatic modeling tools, and developer utilities for 3D design automation.
2. Core Value Proposition: AgentCAD exists to give your coding agent the ability to design real, manufacturable parts. It bridges the gap between text-based code generation and physical 3D output by providing a tight feedback loop where an AI agent can write CAD scripts, automatically validate the geometry, render previews, and iteratively fix errors—producing ready-to-print or ready-to-manufacture files (STEP, STL, GLB) without human intervention in the debugging process.

Main Features

1. Integrated Feedback Loop for Agent Self-Correction: This is the core technical differentiator. When an AI agent (like Claude Code or Codex) writes a build123d or CadQuery script, AgentCAD executes it and returns a detailed metrics object (e.g., dimensions, volume, face_count, is_valid). It simultaneously generates PNG renders from multiple angles. The agent reads both the quantitative metrics and visual renders to catch broken code, confirm the geometry is watertight, and verify dimensional correctness. This creates an autonomous "run, render, inspect, fix" cycle that converges on valid geometry before the output reaches the user.
2. Zero-Boilerplate, Pre-Injected Environment: AgentCAD scripts require no imports. When executed, essential components like build123d primitives (Box, Cylinder, etc.), the show_object function for rendering, and custom AgentCAD edit helpers are pre-injected into the script's namespace. This allows agents to generate concise, focused modeling code without cluttering scripts with setup boilerplate.
3. MCP Server & CLI for Seamless Agent Integration: It operates as an MCP server, making it natively compatible with AI agent platforms like Claude Code, Cursor, and Windsurf. Configuration is simple via .mcp.json. Alternatively, it provides a powerful CLI for standalone use. Installation is streamlined through pip or uv with the [mcp] extra, and it supports a pre-configured "agent skill" installable via skills.sh or ClawHub for one-click setup in agent workflows.
4. Multi-Format Export and Topology Inspection: Beyond rendering, AgentCAD serves as a complete export pipeline. It generates versioned STEP files for CAD interoperability, STL files for 3D printing, and GLB/OBJ files for web viewers and visualization. Its validate command performs pre-execution checks in under 100ms, and its inspect command provides a detailed topology report, allowing both agents and users to debug geometry issues systematically.

Problems Solved

1. Pain Point: The primary problem is the breakdown in AI-generated CAD workflows. AI coding agents can write scripts but cannot verify the physical validity of the output, leading to non-manifold geometry, incorrect dimensions, and non-printable files. This results in endless manual debugging by the user, stifling productivity.
2. Target Audience: This product is essential for AI-enhanced mechanical engineers, programmatic CAD developers, 3D printing enthusiasts using AI agents, and technical designers who want to leverage LLMs like Claude for rapid prototyping. It is for any user of coding agents (Cursor, Claude Code) who needs to generate real-world 3D objects.
3. Use Cases: Key scenarios include automated design iteration from a single prompt (e.g., "design a snap-fit enclosure"), rapid prototyping of mechanical parts (brackets, enclosures, custom tools), educational demonstrations of AI-aided engineering, and batch generation of 3D assets for manufacturing or digital use.

Unique Advantages

1. Differentiation from Traditional Methods: Unlike traditional CAD software (SolidWorks, Fusion 360) which relies on manual GUI interaction, or basic script execution which lacks validation, AgentCAD closes the loop between code and physical reality for an AI agent. It moves beyond simple "code-to-file" conversion to implement a "code-validate-render-fix" cycle that mimics and automates the iterative process a human designer would perform.
2. Key Innovation: The specific innovation is the agent-facing feedback system. The combination of instant geometric metrics (is_valid, volume), automated multi-angle rendering, and a toolset designed to be consumed by an LLM (not a human) creates a unique platform for autonomous, self-correcting AI-driven CAD. It treats the AI agent as the primary user, providing the structured data it needs to improve its own output.

Frequently Asked Questions (FAQ)

1. How does agentcad integrate with Claude Code or Cursor for AI CAD design? AgentCAD functions as an MCP server. You add its configuration to your agent's .mcp.json file. This allows the AI agent to directly invoke AgentCAD commands to execute scripts, validate geometry, and receive results. The recommended quickstart is to have your agent run uvx --from 'agentcad[mcp]' python -m agentcad.mcp on first use, which caches the environment automatically.
2. What file formats can agentcad export for 3D printing and CAD? AgentCAD can export manufacturing-ready STEP files (for interchange with other CAD software), STL files (the standard for 3D printing slicing), and GLB/OBJ files (for 3D web viewers and visualization). The primary output from the agent loop is typically a versioned STEP file along with PNG renders.
3. Does agentcad require specific CAD knowledge or only Python skills? The tool is designed to be used by an AI coding agent. The end-user (you) provides a natural language prompt. The AI agent requires knowledge of the build123d or CadQuery Python libraries to write the modeling scripts, but AgentCAD handles the execution, validation, and feedback. Users can copy example prompts (e.g., for a phone stand or vase) directly into their agent to get started without deep CAD expertise.
4. What are the system requirements and supported platforms for agentcad? AgentCAD is a Python package (supporting Python 3.10–3.12) and runs locally on Linux, macOS, and Windows. It requires no signup and is free and open-source under the Apache-2.0 license. The core dependency is a Python environment with pip or uv for installation.