EarWitness
A local-first Otter alternative, being built live. These are the requirements.
EarWitness transcribes, segments, and diarizes speakers entirely on your machine. An audio tap records the meetings you're in without a bot in the participant list; a bot can join the ones you're not. An Otter-style editor fixes speakers and prose, and MCP tools let Claude work your conversation data. The harness is building it live: the 10 stories below, with their Three Amigos rules and given-when-then scenarios, are the real records driving the build, published before the code exists. The teardown lands here when it ships.
What is EarWitness?
Everything local: the audio, the models, the transcripts, the search index. Nothing leaves your machine. The persona driving the requirements is Dana the Hearing Documenter: someone whose work depends on knowing who said what, who can't hand recordings to a cloud service, and who needs a transcriber working minutes after install.
Audio Tap
Captures system audio so your meetings record without a bot in the room
Local Models
Transcription, segmentation, and diarization run on your machine
MCP
Claude and friends get tools to search and work your conversation data
Meeting Bot
For the meetings you can't attend, a bot joins and transcribes
Editor
Otter-style: fix speaker labels, clean up prose, keep the record straight
Desktop
Elixir LiveView desktop app with a local SQLite store
The Story Map
- Transcribe hearing recordings locally, no cloud, no upload
- Capture your own meetings via audio tap, without a bot interfering
- Speaker diarization: know who said what
- Fix the transcript like Otter: speakers, prose, structure
- Search everything ever said
- Recordings organized and durable
- A working transcriber minutes after install
- Recording-law awareness built into the product
- Your AI assistant works your conversations over MCP
- Send a bot to the meetings you can't attend
The Requirements, Before the Code
Every story below carries its persona, its Three Amigos rules, full given-when-then scenarios, and its acceptance criteria, exported straight from the harness database while the build runs. When implementation starts, these become the BDD specs; when specs pass, browser QA verifies each story against the running app. Watch the statuses change as the build progresses.
Open Source Repository
EarWitness is open source and being built in the open. Releases land as the build progresses: download, break it, and report back. If you have requirements you want in a local transcription app, surface them in the thread and they'll be considered for the story map.
View on GitHubThe teardown lands here
Every completed harness build gets the honest treatment: The Good, The Bad, and The Ugly, with the measured human-vs-agent character split. Cleaner CRM measured 0.4%, Legwork 2.1%, Keel 2.9%. EarWitness gets its number when it ships. Until then, the build updates live on Reddit.