What is buerli-ai?

This package adds an AI chat panel to web applications built with the buerli/ClassCAD 3D modeling framework. Instead of clicking through menus to create or modify geometry, you type a description in plain English and the AI works out which CAD operations to run. All operations execute directly in the browser, no extra server is needed for the geometry work. The package connects to any language model that supports tool calling. Built-in adapters cover Anthropic's Claude API, OpenAI-compatible endpoints including Azure, and local models running through tools like Ollama or LM Studio. A discovery adapter can inspect a given API endpoint to find available models and route requests to the right API surface. Smaller local models may not follow tool-calling instructions reliably, larger models such as GPT-5 or Claude work best. In a React application, setup takes two function calls at startup and two components added to the layout: a panel for the chat interface and an invisible canvas component that lets the agent see the 3D viewport. The panel includes a model picker, a reasoning-effort control, support for image and CAD file attachments, a stop button to cancel a running request, and a code panel that shows the session as a runnable buerli script. The agent can call any of the 254 ClassCAD API methods, read and search the 3D scene tree, get or change the user's current selection, take viewport screenshots, load attached CAD files, and export geometry in STEP, STL, or other formats. You can add domain-specific context to the system prompt to describe your application's naming conventions, units, and common operations, which cuts down the number of back-and-forth turns the agent needs. For production use, the README advises routing AI requests through your own backend so that API keys are never exposed in browser code.