Enable Auto-Attach for MCP Prompts / Embedded Resources

[pattobrien]

VSCode's chat allows for auto-attaching prompt/instructions files, given frontmatter that matches the format expected in the VSCode docs.

It would be really awesome if there was a way to have MCP Prompts / Resources also auto-included in chat context, based on similar frontmatter in either the prompt/resource TextContent itself, or in metadata of the base MCP Prompt or Resource objects.

This would greatly improve the UX of Prompts for end-users, which as of now requires explicit Prompt selection in order to include MCPServer-specific system prompts.

I'm unsure of whether this is functionality that would be best defined in the MCP spec itself, or specifically with the VSCode implementation; therefore, I opened a related discussion in the MCP repo.

[connor4312]

Tailwind folks mentioned that this would be useful for them too: https://x.com/petersuhm/status/1935304863816073441

Martin I think you're taking over prompts? Let me know if you want to sync up here!

[pattobrien]

Tailwind folks mentioned that this would be useful for them too

Sweet! Peter's use case seems identical.

After thinking about it more, this is definitely a Resource feature rather than a Prompt one. And while I think that auto-attach could be standardized by MCP, I feel that the more practical solution for the VSCode client, at the moment, would actually be for the client to treat a file-based Resource (i.e. "file://...") no different from any other VSCode file.

In this particular use case, as long as a Resource meets the same requirements as any other workspace/user-scoped .prompt.md or .instructions.md file, then it should get picked up by the client all the same.

@connor4312 is this something you're considering?

[digitarald]

As a start, having a setting to pin resources as custom instructions might be good.

I agree that the MCP spec should provide some affordance for linking resources or prompts as always or optionally included custom instructions. Otherwise the only way to get into the LLM context is tools; adding up tokens and turns.

[connor4312]

In this particular use case, as long as a Resource meets the same requirements as any other workspace/user-scoped .prompt.md or .instructions.md file, then it should get picked up by the client all the same.

We would not want to auto-discover resources like this necessarily. There is not a way on MCP to "search" resources so we would need to enumerate all resources on all servers for every chat request which is not feasible. But I agree that if we see a resource has a .prompt.md or .instructions.md ending a "pin" action to treat them as persistent custom instructions would be cool.