Better handling of MCP tool ResourceLinks

[fennb]

Description

Currently, MCP Resource Links returned in MCP tool results are transparently downloaded by the client and included in the tool response directly, despite Embedded Resources being a different concept. This could issues in many cases as described below.

From: #3024 (comment) (@fennb):

ie: Imagine a hypothetical tool called find_relevant_documents(query: str, limit: int = 50) that returned 50 resource links to documents (each linked document being 1MB) so that a user can click on the one they want to load/select it/whatever. At the moment, Pydantic AI would try to fetch all the docs immediately and include them in context transparently, which seems... wrong.

An approach would be to leave them as links and then it's up to the user to prompt the agent appropriately in telling it what to do with them (ie: display them verbatim for the actual client to handle or whatever). This is less controversial if the ResourceLink.uri is https://some.server.com/path/file.ext but weirder if it's document://legal/archive/blah.txt (or whatever). I don't really know what an agent would do in that case. Arguably, ResourceLinks maybe shouldn't be used this way, but I think blindly fetching them may not be the right default?

Reply from @DouweM:

I agree the resource link behavior is probably incorrect. I tried to get clarification on that in modelcontextprotocol/modelcontextprotocol#872 (reply in thread), but didn't hear back. It would be easier if the model had a read_resource tool available to it, so it can choose which to load, but that'd need to be provided by the user. Maybe MCPServer should have a flag to include that as a tool, and if provided, we wouldn't auto-read resource links?

Once #3024 is completed, users will have the ability to fetch resources by URI by using MCPServer.read_resource(), which could be exposed to agents by users as a read_resource tool, however, this causes some challenges, as it won't necessarily be clear to the agent/user which MCP server a given resource was returned by (if they have more than 1 MCP server with resources) and they could have overlapping URI schemes (ie: file:// or whatever).

This could potentially be solved by exposing all Resource Links to the agent as (uri, mcp_server_id) tuples and a native/injected read_resource(uri, mcp_server_id) tool that knows how to route the requests appropriately to the correct MCP server.

References

• Support for MCP resources #1783
• Feature: Minimal MCP client Resources support #3024

[pydanticai-bot]

PydanticAI Github Bot Found 1 issues similar to this one:

1. "Support for MCP resources #1783" (95% similar)

[DouweM]

@fennb Good point on multiple MCP servers with resources. Exposing a single read_resource tool for all MCPServers would be a bit tricky to implement, as tools need to be provided by a toolset, and each server is a toolset, but there's no toolset representing all MCP servers. I think we can have a read_resource tool on each MCPServer, which would use the tool_prefix or id as a prefix for disambiguation, then mention in the tool description which MCP server it corresponds to, and also mention that in the ResourceLink JSON representation.