The MCP server is actively being extended, and what gets built next is driven by what builders and their users actually hit. There are two ways to tell us, and they're for different things.
The MCP server exposes a submit_feedback tool that the model calls itself. It needs no OAuth scope, touches no account data, and is the fastest route from "a model couldn't do the thing" to us knowing about it.
It's designed to fire when a task can't be completed as asked — a capability that doesn't exist, a backend call that failed, missing context, or a disallowed action. A model that explains the limitation or offers a workaround is still expected to log it; the explanation isn't a substitute for the report. It's also worth calling when nothing went wrong but there's an observation worth passing on.
Each report carries what the user was trying to achieve and what was tried. It does not ask for or track any Personally Identifiable Information (PII).
For anything that isn't a single failed call — a capability you need on the roadmap, feedback on these docs, or a problem you can't reproduce cleanly — tell us here 💬.
If you're building a connector that works with your users' Typeform accounts, let us know that too — it's how we know who to warn before anything changes underneath you.