Doesn't appear that chatbot knows how to use MCP resources and prompts

[gcmcnutt]

I have noticed a different behavior when compared to, say, Claude desktop, on the ability to walk through MCP resources. In my case, these are essential as resources are metadata descriptions of graphql data elements behind the scenes. As a result, a related query inside this tool can dig into the details, specified in resources, of how to form the query, which fields to use, etc.

Is anyone working on this yet? I think this is primarily an extension to the MCP system prompt -- although this might be a bit more dynamic and templated based on a conversation and tools used (especially to keep system prompt token lengths smaller).

[linear]

MCP-33 Doesn't appear that chatbot knows how to use MCP resources and prompts

[ronaldoguedess]

What do you mean?
Think before calling the MCP? Or call the MCP in the middle of the answer?

[gcmcnutt]

Hi. I've prototyped some ideas where resources are called when referenced from 'prompt' for a particular tool. Basically, context control. The idea is that the resources are both dynamic and potentially large. So, something like:

• a bunch of tools are registered, and their prompts are included as a prefix for the initial user chat
• the model return winds up selecting a tool
• before calling the tool, grab the additional resources if they were referenced (presumably in the tool's prompt) and includes them in a second round trip to the model
• then that reply gets properly synthesized which calls the tool

[a tool prompt]
use me to do X. and if you decide to do this call, grab resources://stuff_for_X and include it for instructions as to how to synthesize the graphql filter

resources://stuff_for_X()
{ return the current metadata for that call based on that user and return a model that is used for synthesizing query }

The above is to help keep context small, but seems troublesome.

The simpler but larger context approach is to:

• at time of a chat
• potentially refresh and read the resources
• potentially replate their references in prompt
• and send the wad down to the model
• assume model will correctly synthesize tool args

I continue to experiment, perhaps others have ideas here. Actually does MCP point to expected uses?

[cgoing-bot]

This is a comment written by the project bot. The content may be incorrect.Thank you for the detailed discussion on MCP resources and prompts implementation.

I appreciate @gcmcnutt's thoughtful exploration of different approaches for context management. Your two proposed strategies highlight an important trade-off in MCP implementation:

1. Dynamic/Lazy Loading: Load resources only when referenced by tool prompts
2. Upfront Loading: Load all resources at chat initialization

Based on your feedback and our current project status, I'd like to clarify our approach:

Current Implementation:

• We support adding custom instructions per MCP server/tool
• These instructions are included in the system prompt when the server/tool is used
• This provides some of the functionality that prompts/resources would offer

Regarding MCP Prompts & Resources:
I admit I don't fully understand the specific use cases and advantages of MCP server prompts and resources, which is why we haven't implemented these features yet.

Questions for the Community:

1. Are prompts typically used as general descriptions of the MCP server's capabilities?
2. Could you recommend example MCP servers that effectively use prompts and resources?
3. What are the main advantages of using MCP prompts over custom system instructions?

For prompts, I believe we could implement periodic validation and caching when connecting to MCP servers. However, resources require more consideration due to their potentially dynamic nature.

The flow you mentioned: prompt -> resource (optional) -> tool makes sense for better LLM understanding of MCP servers. I'd love to see some concrete examples to better understand the implementation patterns.

Would anyone be willing to share example MCP servers or documentation that demonstrate best practices for prompts and resources?

[cgoinglove]

No plans.

[brrock]

TLDR we only support tools right now