How to use MCP resources with agents and LLMs?

[MadanMaram]

Hi, I’m exploring MCP resources and how they can be integrated with agents and LLMs. I have a few questions and would appreciate guidance or examples:

1. Resource selection similar to tools:
In tools, we have a list of tool names and descriptions, and the agent/LLM can pick the right tool for a query.

Can we implement a similar mechanism for MCP resources, where a list of 3–4 resources is evaluated and the LLM chooses the most relevant one?

What metadata or description exists for MCP resources that can help an agent/LLM make this selection?

2. Resource usage as tools:

Is it possible to “use” resources in the same way as tools, where the agent dynamically selects and invokes the resource?

If yes, are there examples or a recommended approach for proving this works in an implementation?

3. Agent integration:

How does an agent work with MCP resources?

If the agent cannot directly consume resources in a standardized way, how can MCP resources still be useful to the agent or LLM for answering queries?

4. Implementation guidance:

Any code samples or references for dynamically selecting resources based on a query using an LLM or agent would be very helpful.

Thank you!

@jlowin I’m exploring MCP resources and their integration with agents/LLMs. I have a few questions about usage of resources. I’d really appreciate any guidance, examples, or pointers you could share. Thank you for your time!

[jlowin]

Generally, in contemporary clients resources are left up to human users to add to context. However, a more interesting practice is to introduce tools to accomplish what you described. This lets the agent make decisions about resources independently.