// a weekend project · built with Claude Code
make reading
interesting again.
Articles and books, read aloud by a neural voice — entirely on your Mac.
Ensuring Safety in the Generative AI Ecosystem
space play·←/→ ±5s·t transcript·tap any word to jump there
A real Summary-mode read — local LLM + CSM-1B, in codeword, a custom-tuned clone voice.
// concept
Nothing is live, so quality wins
Every read is synthesized up front rather than streamed. There's no live turn to keep pace with, so there's no underrun and no echo — and voice quality gets to win over latency.
// walkthrough
The five things you'll actually do
Start it
readback-cli, from anywhere. It boots the server itself, and opens on the newest posts from the blogs you follow.
Pick something
Press 1–5 for one of those picks, or paste your own link or image path.
Drive the player
space pauses, ← and → jump five seconds, t toggles the transcript.
Change your mind
/voice, /model and /mode swap mid-session — the next read picks it up, with no restart.
Leave
q quits, and takes the server down with it.
// how it works
Local model, real player, a Pi if you want one
One server, two clients: extract → optional summary → synthesize offline with CSM-1B, once — then a separate, model-free path just lists the library and replays a finished WAV. Generate once, replay many.
offline
[100% on-device]Nothing is uploaded and nothing is billed per character. Fetching a URL is the only thing here that touches the network.
sources
[url + image + scans]A URL, an image, or a folder of page photos — one local vision model OCRs them into a continuous read.
voice
[csm-1b]Neural TTS from Sesame; clone any voice from a short clip, or LoRA fine-tune.
modes
[full / summary]Read it verbatim, or let a local MLX LLM rewrite it as a spoken explanation first.
player
[seek + transcript]A transcript that highlights word by word as it reads — in your shell, or in the browser.
library + network
[sqlite + pi]Every read saved and searchable; deploy the dashboard to a Raspberry Pi via PiZoW and replay from any device at home.
6 features · one local process · nothing uploaded