Support MCP elicitations both in client and server

[valdo404]

Description

As an AI programmer, in order to better manage inputs and outputs of a tool, would it be possible to support MCP elicitations ?

MCP specification describe this as a way to exchange information between an mcp client and an mcp server:

https://modelcontextprotocol.io/specification/draft/client/elicitation

do you think this could be implemented into pydantic AI ? can an agent and a tool talk by using this api ?

best regards

References

No response

[AntonGomes]

See #2258

[DouweM]

@valdo404 Thanks for filing this issue, that's definitely a feature we'll want to support. Let's discuss what that could look like.

The simplest thing we can do right away is to let users specify an elicitation_callback on the MCPServer and handle parsing the params and building the response themselves, using the types from the mcp package. That's already implemented in that PR that was linked, among other things.

If you're imagining some more native or opinionated integration with the agent, I'd love to hear what you have in mind. One hurdle is that the elicitation request message is arbitrary text that is expected to be presented to the end user who then provides a response (or not), while Pydantic AI is agnostic as to how users are brought into the loop.

The AG-UI integration may offer a way to natively support this, but it's not clear yet how the AG-UI protocol as defined could support elicitation; I've filed an issue on their side: ag-ui-protocol/ag-ui#231.

[valdo404]

One thing that could be interesting, is a father / child integration type, where a child agent is asking for informations to its parent agent. I'm thinking about autonomous multi-agent use cases that would enable bidirectional integrations between agents.

Some kind of UI toolkit would be appreciated too. And I guess ad-ui is filling the gap here.