earshot: hear your agent's last reply, action first
Turns the last Claude Code reply into a short spoken summary that leads with what you need to do, then the result. The model stays local.
About it
earshot speaks the last Claude Code reply as a short briefing: what you need to do, then what happened. earshot say reads any text aloud as written. --dry previews with no audio. In Claude Code the summary is /. and raw narration is /say. The summary model is local Ollama. The README's example was worded by gemma4:26b; the default model is llama3.2:3b. Limits: It needs python3, jq, curl, Ollama and a local model. The installer offers to pull llama3.2:3b. The default source is the current Claude Code transcript. Adapters for other agents are not built; pipe the text in. The system voice is macOS say or Linux espeak-ng. A neural voice needs a TTS server, and ffmpeg is only for that path. Pointing the model or voice at another machine sends that call there. Door 31 of 100 Doors.