I'm new to local AI agents and want to experiment with one that can run entirely on my own hardware. The codebase is made up mostly of VB.NET client applications, with an MSSQL project included in the solution. Ideally, the agent would inspect every project and produce Markdown documentation explaining how the applications work from a user's perspective—for example, describing that clicking a green button displays a message box about the selected object.
I tried a cloud-based coding model, and the results were surprisingly good, but I can't use it on the real codebase because the source is confidential. I have an M3 Max available. What local models and tools would be a good starting point for this workflow?
2 Answers
Qwen 2.5 Coder 32B is worth trying locally for this. You can run it through Ollama and connect it to an editor integration such as Continue. Point the agent at each project directory, give it a system prompt explaining that it should describe user actions by following event handlers, and have it produce Markdown files as it works through the code. A larger model will be slower, but it should handle VB.NET reasonably well on capable hardware.
A recent Qwen coding model should also be capable of this kind of documentation task. Ollama is one of the simpler ways to download and serve a model locally, but the important part is the workflow: process one project or a small group of files at a time, ask the model to trace event handlers and related methods, and save the result as Markdown. Review the generated documentation afterward, since UI behavior can be spread across multiple files and the model may miss database-side logic.
Do I need Ollama for this, or is there another setup? I’m completely new to running models locally.