Making a self-serve buying flow from an MCP connector is relatively nightmarish right now. (1) we can't conditionally and deterministically show any kind of buy button and (2) we can't see what the LLM tells the user. It is extremely painful. More generally error handling is like shooting in the dark.
An MCPApp doesn't solve this because views are linked to a tool call, and since there is no conditional triggering of the MCPApp, well, it's bad.
cc @mxstbr @TobinSouth
I actually really like having only one openai dot instead of many grok bots. Removes the cognitive overload of having to choose which bot to talk to. I wouldn't mind if this dot then created its own dots, but I only want to speak to one.
LLMs are so focused on not hallucinating or making mistakes that they are now absolutely incapable of writing even a paragraph that reads well. They can at best just clean/polish a text that's already 95% there.
Making a good MCP and AI plugin is hard. OAuth flows are not trivial to get right, especially with multiple accounts. Getting your tools to trigger at the right time and writing good descriptions is not easy at all. Skills get truncated randomly in Claude. (and AI is extremely bad at writing skills and descriptions - don't write your tool descriptions and skills with AI).
The fact that it's so easy to fix small but super annoying IT problems like this now is so awesome. Thank you Codex.
Is there someone @claudeai @ClaudeDevs who I could contact to get help with getting our MCP verified? Got no answer for weeks from the directory email.
I could swear 6 Pro is worse than 5.6 Pro. 5.6 Pro had bouts of brilliance whereas 6 Pro is very average, erring on the sloppys slide. Am I hallucinating?
Feature request in Google Docs: token counter. cc @ChanduThota @GoogleWorkspace