Feature Request: Custom prompts

[Ayfel]

It would be very useful if we could inject a custom prompt, possibly by adding it before starting the server. If its already a feature I might have missed it.

Thanks.

[Scriptwonder]

Can you elaborate more on what you mean by "custom prompt"? The closet one I can think of is custom tool that the server can call (after set and rebooting)

[Ayfel]

Its a way to customize the instructions the server provides the LLM on successful connection/handshake. This way we can provide a context to the LLM before even calling a single tool, and give instructions on how to best utilize these tools, order, etc. It is part of the MCP spec I believe.

Some references below:

https://blog.modelcontextprotocol.io/posts/2025-11-03-using-server-instructions/

https://modelcontextprotocol.io/specification/draft/schema#initializeresult

[Scriptwonder]

A very good suggestion, I think. Right now, we don't offer a custom prompt because we dump all the tool descriptions into LLMs when starting the server. As the repo grows and more features emerge, an adaptive instruction/description set might be beneficial for long-term development (customization and efficiency). Not sure how to implement it yet, given how our server is structured right now, but we will look into it further. CC'ing @msanatan @dsarno to broaden the discussion.

[msanatan]

I knew we had instructions before, I can't recall why it was removed. Likely because we didn't maintain it. We have some instructions in main that's pretty much to tell agents how to access custom tools in a multi-tenant environment.

The biggest issue with instructions back then is that it's a feature that only Claude Code supported. Maybe that's changed since?

And given that "Skills" is a feature with broad support, maybe we invest in that to guide agents?