You say thousands of sentences a day. How many do you remember?

You promise something in a meeting and forget it by the elevator. A friend mentions a number or a time and two days later it’s gone. An idea comes up again and again and never gets written down.

The existing options either want you to buy a recorder or upload all your audio to someone else’s server.

TaoSheng’s idea is simple: your phone is always with you, and it already has a microphone. Let it listen quietly, record only when someone speaks, transcribe in batches, work out who said what, and let an LLM turn it into things you can find later.

An app written by AI

Almost all of TaoSheng’s code was written by AI: Claude on architecture, interfaces and every screen; Codex on capture, transcription, storage, push and backup. The human files requests, tests on a real phone and says what’s wrong. Nearly two hundred task cards and delivery notes are public in the repository — mistakes included.

Claude · architecture & UI Codex · internals Human · testing & taste
Read the task cards →

Principles

On-device first

Voice detection, voiceprints and local transcription run on the phone.

No lock-in

Swap transcription and LLM providers any time, including self-hosted ones.

Your data

No account, no server of ours, raw audio deleted after 3 days by default.

In the open

Source, task cards and issue logs are public. Audit it, build it yourself.

Thanks

Thanks to everyone who contributed code, and to feitian124 for building the first version of this website.

License

From 0.4.0 the source is available under FSL-1.1-MIT: personal, educational, research and internal company use are unaffected; competing commercial products or services are not allowed. Each release becomes MIT two years after it ships. Versions up to 0.3.1 remain MIT.

LICENSE.md →