Live Ozan Radio

Personal AI radio — no catalog music, ever. DeepSeek is the DJ. Google Lyria 3 composes every track. Taste comes from settings.json (and optionally Spotify).

Inspired by Magenta RealTime 2 (live, ~200ms) and Lyria 3 (full songs via Gemini API). On Mac you can layer MRT2 for true live improvisation; this repo ships the cross-platform Lyria + DeepSeek stack first.

Stack

Layer	Product	Role
DJ brain	DeepSeek (Tinqs proxy or BYOK)	Mood, prompts, chat, variety
Music engine	Google Lyria 3 Pro / Clip	Generate MP3 tracks
Taste	settings.json + taste_seeds.json	Genres, mood, instruments — DJ reads every request
Taste (optional)	Spotify Web API	Top artists, genres — never plays Spotify
Player	FastAPI + gateway/player.html	Stream queue, library, chat, dashboard
Live (optional)	Magenta RealTime 2	Apple Silicon only — see below

Player dashboard

Open http://127.0.0.1:8787/player after starting the server.

Feature	What it does
Cost dashboard	Today's estimated spend, per-track cost, songs generated vs daily cap
Settings (gear icon)	Shuffle mode, mix saved + new tracks, new-song chance, max songs/day
Shuffle mode	On track end or Skip, picks a random saved song or composes a new one
Saved songs	Click any track in the library to play
DJ chat	Talk to DeepSeek — requests can trigger new Lyria generations

Shuffle behaviour

When shuffle is on (default):

1. If you have saved songs and daily quota remains, each next track has a new_song_chance (default 35%) of being freshly composed.
2. Otherwise a random saved song plays (never the same track twice in a row if alternatives exist).
3. If the library is empty, it generates until the daily cap is hit.

Daily generation stats live in songs/stats.json (gitignored, local runtime only).

Taste (settings.json)

Edit settings.json at the repo root. The DJ reloads it on every generate and chat.

{
  "taste": { "summary": "...", "genres": [], "mood": [], "instruments": [] },
  "playback": {
    "shuffle": true,
    "mix_existing_and_new": true,
    "new_song_chance": 0.35
  },
  "limits": { "max_new_songs_per_day": 10 },
  "costs": {
    "lyria_pro_usd": 0.08,
    "lyria_clip_usd": 0.04,
    "deepseek_per_track_usd": 0.002
  }
}

Default taste profile: ethnic world dubtronica (global roots + dub space + electronic groove).

The player settings panel PATCHes playback and limits via /api/settings and writes back to this file.

Saved songs

Every track is written to ./songs/ and committed via Git LFS (audio) + plain git (metadata):

File	Storage	Contents
{id}_{title}.mp3	LFS	Audio
{id}.meta.json	git	Title, mood, DJ line, prompt, lyrics, timestamp
manifest.json	gitignored	Runtime queue index

Browse in the player under Saved songs, or GET /api/songs. After clone: git lfs install then git lfs pull.

Quick start (Forge / Windows)

cd live-ozan-radio
python -m venv .venv
.venv\Scripts\activate
pip install -e .
copy .env.example .env
# Fill GEMINI_API_KEY + DEEPSEEK_API_KEY (and Spotify if you have them)

$env:DEEPSEEK_BASE_URL = "https://api.deepseek.com/v1"
python -m ozan_radio serve
# Open http://127.0.0.1:8787/player

One-shot track (no server):

python -m ozan_radio generate

Environment

Variable	Required	Notes
GEMINI_API_KEY	Yes	Google AI Studio — Lyria 3
DEEPSEEK_API_KEY	Yes	Tinqs proxy token or DeepSeek direct
DEEPSEEK_BASE_URL	No	Default https://api.deepseek.com/v1
SPOTIFY_*	No	Refresh token flow — taste only
LYRIA_MODEL	No	lyria-3-pro-preview (default) or lyria-3-clip-preview
RADIO_OUTPUT_DIR	No	Default ./songs

Cost estimates (defaults)

Model	~USD / track
Lyria 3 Pro + DeepSeek	~$0.082
Lyria 3 Clip	~$0.04

At the default cap of 10 new songs/day with Lyria Pro, projected max spend is ~$0.82/day. Adjust costs and limits in settings.json or the player settings panel.

Spotify setup (optional taste)

1. Create an app at Spotify Developer Dashboard.
2. Add redirect URI http://127.0.0.1:8888/callback.
3. Complete OAuth once to obtain a refresh token (scope: user-top-read).
4. Paste SPOTIFY_CLIENT_ID, SPOTIFY_CLIENT_SECRET, SPOTIFY_REFRESH_TOKEN into .env.

If Spotify is not configured, the DJ uses settings.json + taste_seeds.json.

API

Method	Path	Description
GET	/api/now	Current track metadata
GET	/api/queue	Full queue
GET	/api/stats	Dashboard: today's spend, quota, cost estimates
GET	/api/settings	Playback, limits, costs
PATCH	/api/settings	Update shuffle, daily cap, etc.
POST	/api/generate	DJ plans + Lyria renders next track
POST	/api/shuffle/next	Smart next: library shuffle or new generation
POST	/api/skip	Advance (uses shuffle when enabled)
GET/POST	/api/chat	DJ chat log and messages
GET	/api/songs	Saved library
POST	/api/songs/{id}/play	Play a saved track
GET	/stream/{file}	MP3 stream
GET	/player	Web UI

Magenta RealTime 2 (optional live layer)

On Apple Silicon (Kraken), install magenta-rt for sub-second live generation:

uv pip install "magenta-rt[mlx]"
mrt models init && mrt models download
mrt mlx generate --prompt "disco funk" --duration 4.0 --model=mrt2_small

Wire MRT2 as a bridge between tracks or as a live “bed” under the Lyria queue — PRs welcome.

Repo

Public on Git Studio: https://tinqs.com/tinqs/live-radio

git clone git@ssh.tinqs.com:tinqs/live-radio.git
cd live-radio
git lfs install && git lfs pull

Static player preview: https://tinqs.com/tinqs/live-radio/src/branch/main/gateway/player.html (shell only — audio streams from your running server).

Agent usage

DeepSeek (Pi, Cursor, Claude Code) can operate the station via HTTP or the skill in .cursor/skills/ozan-radio/SKILL.md.

License

Apache 2.0 — same spirit as Magenta RealTime 2 open weights.