Earshot — your agent can hear now

An audio editor your AI agent can hear. Via 43 WebMCP tools it detects silences, reads the transcript, and proposes surgical edits on the timeline — you approve.

Comment

Inspiration

A while back I really wanted to be a YouTuber. The part that always killed it for me was editing — cutting out the ums, the long pauses, and the takes where I fumbled.

It was slow and boring, and I'd give up. Now I know there are agentic tools for video editing but I did not find anything for just audio editing.

What it does

Earshot is a browser audio editor where your AI agent can actually hear the recording.

It exposes its tools over WebMCP, so an agent can:

  • Find silences, filler words, and long pauses
  • Read a word-level transcript
  • Cut audio just by removing words from the transcript
  • Tighten pauses, level a quiet speaker, and add chapters
  • Export the finished audio and captions

The agent suggests the edits, and then I can approve

How we built it

Built with Codex and Claude doing most of the coding, plus a fair bit of manual product decisions from me on how it should look and feel.

It runs entirely in the browser, with Whisper for transcription either through OpenAI or locally on your GPU.

Challenges we ran into

  • It was honestly pretty simple, codex and claude helped a lot. The only thing I had to work on was to make sure the transcription editing for accurate with whisper running on GPU

Accomplishments that we're proud of

  • The whole loop works end to end
  • Editing audio by editing text actually feels natural
  • Transcription can run fully on-device, no key needed

What we learned

  • The best part isn't automating everything — it's letting the agent do the boring work while I keep the final say
  • WebMCP will make it fun to use boring products

What's next for Earshot — your agent can hear now

I will keep it opensource

Built With

Share this project:

Updates

Submission history