A prompt becomes a sequence of CAD operations: building an agent loop from scratch to understand tool use.
I built Solid Agent from scratch to understand how an agent connects a language model to real application tools. It is a command-line Python prototype that operates Solidcraft's CAD tools through MCP.
- My role
- Python agent loop & MCP integration
- Context
- Working learning prototype
- System
- Tool calling · CAD · context management
Demonstration
From a prompt to a model
The image below shows the result of the prompt “Build a car in low poly”: a red car in Solidcraft, alongside the agent's terminal and operation history.
What I built
The loop discovers MCP tools, exposes their schemas to a Chat Completions-compatible model, executes requested calls and returns observations to the conversation. The model can then continue working or give a final answer.
I implemented tool validation, blocked commands, operation history and context compaction in Python. Checkpoints combine the current CAD state with earlier successful operations and recent exchanges.
Current state
A working learning prototype for exploring tool use in CAD. It requires Solidcraft and a tool-calling model endpoint; the examples demonstrate the interaction loop, not a general guarantee of modelling accuracy.
