Show HN: AgentVoice – Get feedback from AI to improve your MCP

I wanted to share a tool we built to solve a specific frustration: AI agents (like Claude Code or Cursor) are becoming the primary consumers of our APIs, but they are "silent" users. Unlike human developers who file GitHub issues or post on Discord when a schema is confusing or an error message is cryptic, agents just fail, retry, or give up.

We ran a trace on our own MCP server and found 9 distinct friction points in a single 10-minute session. Things like undocumented enum constraints and case-sensitivity issues that never showed up as "errors" in our standard logs.

AgentVoice closes this loop by triangulating three sources: it runs an LLM observer over session transcripts to diagnose intent, pulls production telemetry to prove the scale of the failure, and adds a submit_feedback tool so agents can narrate their own friction in real-time.

I’d love to get feedback on the approach, especially from anyone currently shipping MCP servers or agent-facing APIs.