Idea: Present Info Being Written To Record/Base Prior To Actually Writing

[jpaulgale]

I assume there's probably something in the works like this as a general purpose tool for MCPs but would be great to be able to see any write changes within the chat interface, prior to their being written and to accept them with an "Accept Changes" button.

Separately, getting a simplified version of the mcp's actions in general (sort of like CoT) might be helpful for spot checking -- could be hidden by default.

ex.

Searching for ({Name} = "John Doe")
Found 1 record
{{full record details in table form with header row}}

{{LLM Response}}

Great work on this project, would be happy to work together on expanding it, if that's of interest.

[domdomegg]

Yup I agree with this! I wrote up something vaguely along these lines at a spec level here: modelcontextprotocol/modelcontextprotocol#114

(especially the bit about "it might return an impact statement [...]")

[domdomegg]

At the moment it's hard to implement this as part of an MCP server - but once something like that gets brought into spec I'd be keen to implement it well here.

[jpaulgale]

Good to know the context -- is there any work being done on the dynamic UI piece in re: MCPs? I know that openai has done some work on it.

[domdomegg]

I'm not aware of anything related to dynamic UI with MCPs at the moment unfortunately, no

[jpaulgale]

Nice implementation here from Raycast (different protocol, but a good example)