Elicitation

Whenever I've tried to use my LLM to create git commits or pull requests or something like that, I hate it's writing style. I frequently find myself spending more time trying to instruct the model properly than I would've required to edit the message myself. The MCP elicitation feature enables servers to request additional information from the users for specific tasks, for instance while executing tools. The ElicitationMiddleware in wags uses this to improve the user experience for such tools where users might want to edit the models generation before executing them directly.

Example

RequiresElicitation type annotation on on tool parameters informs wags middleware which values should be reviewed by users.

from typing import Annotated
from wags.middleware import ElicitationMiddleware, RequiresElicitation

class GithubHandlers:
    async def create_issue(
        self,
        owner: str,
        repo: str,
        title: Annotated[str, RequiresElicitation(
            "Issue Title"
        )],
        body: Annotated[str, RequiresElicitation(
            "Issue Body"
        )]
    ):
        pass

and the ElicitationMiddleware can be enabled in the wags proxy:

main.py

handlers = GithubHandlers()
mcp.add_middleware(ElicitationMiddleware(handlers=handlers))

Now whenever a tool with elicitation annotations is triggered, wags will intercept the tool call, allow the user to edit those arguments, and then make the upstream tool call. So you can be confident that your agent doesn't cause embarrassment for you

API Documentation

Marks a parameter for user review before tool execution.

Parameters:

Name	Type	Description	Default
prompt	str	Message shown to the user describing the parameter	required

Triggers elicitation dialogs for marked parameters.

When tool handlers have parameters with RequiresElicitation annotations, this middleware presents those values to the user for review and potential modification. Aborts execution if the user declines.