I made my coding agent talk Quick context: I use Claude Code and Codex daily and noticed I was spending half my "agent is working" time just sitting there watching the screen. I was like, what if Claude or Codex can just talk back at me, like Jarvis did Ironman, so I don't have to go through all the output soup? So I built Heard. What it does: Speaks your agent's intermediate output - tool calls, status updates, the prose between actions. You can get up, make coffee, and still hear when it hits a failure or needs input. Stack: \- Python daemon, Unix socket, fire-and-forget hooks (never blocks the agent) \- ElevenLabs for cloud TTS, Kokoro for fully local (no key needed) \- Optional Claude Haiku 4.5 for in-character persona rewrites \- Adapters for Claude Code + Codex; \`heard run\` wraps anything else \- macOS app + CLI, Apache 2.0 What I learned building it: The hard part wasn't TTS, it was deciding what NOT to say. First version narrated everything and was unbearable in 90 seconds.…