agentcad

What is agentcad?

agentcad is an open-source CAD tool for coding agents. It provides both an MCP server and a CLI so an agent can generate, render, validate, and export real CAD output from Python-based workflows.

The product is designed for human-agent collaboration: it gives the agent a feedback loop with geometry metrics, previews, and validation results so it can inspect its own work and correct mistakes before handing designs back to the user. According to the site, it runs locally, is free, requires no signup, and supports STEP, STL, GLB, and OBJ workflows.

Key Features

• Open-source MCP server and CLI: supports agent-driven CAD workflows through either local command-line use or MCP integration in editors and agent environments.
• Render-and-validate loop: returns a render plus metrics such as dimensions, volume, face count, and validity so the agent can check geometry before finishing.
• Export support for common CAD formats: can output STEP, STL, GLB, and OBJ for downstream use in printing, viewing, or CAD handoff.
• Versioned execution and diffs: produces versioned STEP files and supports comparing versions to track design iteration.
• Pre-execution checks and topology inspection: validates inputs before execution and reports geometry/topology details to help debug issues quickly.
• Local-first setup: runs locally, with install options for pip, uv, and MCP configuration for tools such as Claude Code, Cursor, and Windsurf.

How to Use agentcad

Install the package locally, then connect it to your coding agent either through the CLI or by adding the MCP server configuration to your agent workspace. The site shows Python 3.10–3.12 support, with a pip install 'agentcad[mcp]' path and a CLI-only install option.

Once connected, give the agent a CAD prompt such as a phone stand, enclosure, or vase. The agent can generate geometry, inspect the render and metrics, revise the model, and export the final files when the design is valid.

Use Cases

• Rapid concepting in an AI coding agent: turn a natural-language prompt into a basic CAD model and preview it in the same workflow.
• Printable enclosure design: create simple mechanical parts such as snap-fit boxes, covers, or device enclosures and export them for 3D printing.
• Geometry debugging and iteration: use metrics, validation, and topology reports to diagnose why a model failed or produced bad geometry.
• Web or viewer-ready 3D assets: export GLB or OBJ for lightweight visualization in apps, sites, or asset pipelines.
• Agent-assisted educational or reconstruction projects: build measured reconstructions or structured models where the agent needs repeated feedback to converge on a faithful result.

FAQ

Does agentcad run locally? Yes. The site describes it as running locally, with install options that do not require signup.

What outputs does it support? The site mentions preview renders plus STEP, STL, GLB, and OBJ export.

Does it work with MCP-based agents? Yes. The product includes an MCP server and shows example setup for Claude Code, Cursor, and Windsurf.

What Python versions are supported? The page says agentcad currently supports Python 3.10–3.12.

Is it open source? Yes. The source is listed as Apache-2.0 on GitHub.

Alternatives

• Traditional CAD software such as parametric desktop CAD tools: these are typically used directly by human designers rather than by an AI agent, and they offer manual modeling workflows instead of agent-driven execution.
• Script-first geometry libraries: these suit developers who want to build CAD shapes directly in code, but they may not include the same agent feedback loop, render reporting, or MCP integration.
• 3D modeling tools with export features: these can produce printable or viewable assets, but they are usually centered on manual modeling rather than automated prompt-to-geometry workflows.
• Other agent toolchains for code generation: these may help an agent write code, but they do not necessarily provide CAD-specific validation, preview, and version-diff support.