Case study

Earshot

A coding agent's tool calls turned into sound and speech, so you can follow what it's doing without watching the terminal.

Problem

A coding agent narrates almost nothing by default. It reads a file, edits another, runs a shell command, and if you're not looking at the screen the moment it happens, you've missed it. Working alongside one means either staring at the terminal or finding out after the fact what it touched.

Approach

It exposes a hook for the full tool call lifecycle: reads, edits, writes, shell commands, permission prompts, subagents. I attached a distinct sound to each one, a single sound per call. The closing message at the end of a turn is read aloud through screen reader integration, with a Windows SAPI fallback, and only while the agent has focus. If you miss it, you can replay the last message.

Outcome

Published it as a plugin marketplace entry instead of leaving it as a personal script. It reads nothing over the network and writes nothing I didn't ask it to; the one log that would record what got spoken to me is off by default. People who aren't me use it now, which is the actual bar for whether something like this is finished.

Built with

Links