Support for human-in-the-loop aka elicitations from MCP servers

[owengo]

Currently it is not possible to put a "human in the loop" to validate a tool execution.
But the MCP protocol supports "elicitations", which are messages from a MCP server asking for confirmation , or confirmation + some data before executing a command.
For example a shell access MCP server could ask confirmation before executing dangerous commands, a google drive MCP could ask before suppressing a document etc.

For example in librechat we could have this kind of input generated by the MCP server before executing some actions:

It's also possible to request some data via a form:

The full spec is at https://modelcontextprotocol.io/specification/2025-06-18/client/elicitation

[owengo]

The new Apps SDK by OpenAI indicates ChatGPT supports elicitations as an MCP client: https://developers.openai.com/apps-sdk/build/mcp-server

[Zzebra99]

I am also interested in knowing if there are plans to introduce support for this?

[owengo]

@Zzebra99 I made this PR #8684 if it can help

[dvejsada]

Hope it will get merged soon. Unfortunately, lately a lot of great features are getting stuck at unmerged PRs...