With AI becoming more and more in the development world, my company now has us using Github CoPilot within VS.
With copilot, we can add copilot-instructions.md.
Are there any plans of Radzen to supply an instruction set for use in CoPilot to give the AI a better chance of completing the tasks?
I’ve used both VSCode and Cursor without any specific instructions - worked great.
RBS, Cursor with Claude 4 Sonnet, coding way above my pay grade, in the indexing and Docs, in Cursor I added the links to RBS Components and Documentation, it is indexed. Cursor has access to the documentation, though an RBS MCP Server might be handy. Tried using GitHub Copilot, did not like it.
I've spent the day creating a Radzen MCP server which is fed with the github repository. It somewhat works. One thing bothers me though - the agent (VS Code) often doesn't use the MCP at all and uses the LLM (GPT 4.1) directly. I've checked what other vendors are doing and ours seems to be in line. I found however that using the MCP is not guaranteed at all and sometimes needs changing the prompt. How do you expect the MCP to work?
one problem that Ai has that is annoying, is no matter how many rules and MCP info you give it, it reverts back to regular html, CSS Classes , and javascript, rather than staying with RBS Syntax and CSS classes, though Claude 4 Sonnet seems to work really well without the rules and Mcp servers, but Cursor pricing has gotten more expensive so this is a consideration. my workflow is using Claude to write SQL Scripts for the Database, which i does extremely well, then infer with RBS, and then make a folder and add pages for the module that I'm working on then use Claude to refactor the Layout, Modify the Import and Export of the Data, (Excel, CSV, Pdf, print, and Email Automations). For An MCP Server, and I dont know anything about them, if there is a way to keep the AI using the RBS Syntax, and CSS Classes, and Components, so the Code base is consistant is all I can think of.
I don't think the MCP would alone solve that. As far as I understand the LLM decides on its "own" whether to use an MCP or not. And more often than not it doesn't use the MCP unless the prompt is very specific. I am still not convinced that the MCP actually adds any value to be honest. I will come back to this in case some new information pops up.
These new tools are making a mediocre developer look good, though, you really have to keep your version control locked down, as AI can mess stuff up extremely fast, if you are not paying attention, I learned the hard way after not doing a code review and made some commits that affected the web sockets resetting every minute, took 3 days to figure it out.