hexu

A self-hostable clone of Huxe — the AI audio app shutting down 2026-05-28. Produces a daily personalized podcast briefing from your Gmail, Google Calendar, and RSS feeds, with two-host NotebookLM-style dialogue and an RSS feed any podcast client can subscribe to.

Built on OSS components: Podcastfy for the audio pipeline, edge-tts (free, keyless default) or your own local OmniVoice Studio for voices, LiteLLM for model routing (Gemini 2.5 Pro by default, Claude as fallback), and the standard Google API client for inbox + calendar access.

Quickstart

# 1. install
git clone https://github.com/broomva/hexu && cd hexu
uv sync

# 2. configure (edit values)
cp .env.example .env

# 3. authorize Gmail + Calendar (one-time)
uv run hexu setup-oauth

# 4. run the first briefing
uv run hexu brief

# 5. serve the feed locally + subscribe in any podcast app
uv run hexu serve
# → http://localhost:8080/feed.xml

Architecture

sources → context → dialogue → audio → publish
─────     ───────   ────────   ─────   ───────
Gmail              LiteLLM     TTS     RSS feed
Calendar    →      Gemini 2.5  Sesame  feedgen
RSS feeds          (Claude     Kokoro  Local / S3
                    fallback)  Podcastfy

Each stage is a pluggable layer:

Stage	Module	Default	Alternatives
Email	fetchers.gmail	Gmail API	—
Calendar	fetchers.calendar	Google Calendar API	—
News	fetchers.rss	feedparser	—
LLM	dialogue.generator	gemini/gemini-2.5-pro	any LiteLLM provider
TTS	tts.*	podcastfy	omnivoice (local, recommended), sesame (GPU), kokoro (CPU ONNX)
Feed	feed.publisher	feedgen → iTunes-compatible RSS	minimal RSS fallback
Storage	feed.storage	local	s3 (AWS, R2, MinIO)

Configuration

All settings load from environment variables or .env (see .env.example).

Variable	Required?	Default	Meaning
GEMINI_API_KEY	yes (if Gemini)	""	Google AI Studio key for Gemini 2.5 Pro
ANTHROPIC_API_KEY	optional	None	Claude fallback
GOOGLE_OAUTH_CLIENT_SECRET_PATH	yes	./client_secret.json	Downloaded from Google Cloud Console
GMAIL_LOOKBACK_HOURS	no	24	How far back to scan inbox
CALENDAR_LOOKAHEAD_HOURS	no	24	How far forward to scan calendar
RSS_FEEDS	no	[]	JSON array of feed URLs
TTS_PROVIDER	no	podcastfy	omnivoice / podcastfy / sesame / kokoro
OMNIVOICE_BASE_URL	no	http://127.0.0.1:3900	OmniVoice backend (omnivoice only)
OMNIVOICE_STEPS / _LANGUAGE / _SPEED	no	16 / Auto / 1.0	OmniVoice tuning
VOICE_HOST_A, VOICE_HOST_B	no	provider-specific	Voice IDs
OUTPUT_DIR	no	./output	Audio, manifest, feed.xml
FEED_BASE_URL	no	http://localhost:8080/	Base URL embedded in feed
LLM_MODEL	no	gemini/gemini-2.5-pro	LiteLLM provider/model
STORAGE_BACKEND	no	local	local or s3
S3_BUCKET	yes for S3	None	Bucket name
USER_NAME	no	friend	Threaded into the dialogue

TTS provider selection

omnivoice (recommended — the self-owned keystone) — Points the clone at a local OmniVoice Studio backend (FastAPI on 127.0.0.1:3900). Voice cloning, 646 languages, free at any volume, nothing leaves your machine. This is the functional path that makes the stack fully durable — no rented voice, no per-character billing. Start the backend, then:

# one-time: install the skill helpers + start the backend
npx skills add broomva/omnivoice-skill
bash ~/.claude/skills/omnivoice/scripts/start-backend.sh   # health: GET /system/info

# create voice profiles in the UI (localhost:3901), note the profile_ids, then:
#   TTS_PROVIDER=omnivoice  VOICE_HOST_A=<profile_id_a>  VOICE_HOST_B=<profile_id_b>
uv run hexu brief

If the backend is unreachable the renderer raises an actionable error (or emits a silent placeholder when HEXU_ALLOW_SILENT=1). Tune OMNIVOICE_STEPS (8 draft / 16 balanced / 32 quality), OMNIVOICE_LANGUAGE, OMNIVOICE_SPEED.

podcastfy (zero-setup default) — Works out of the box, no GPU required. Hands the rendered transcript to Podcastfy's built-in pipeline. Selects the backend via PODCASTFY_TTS_MODEL: edge (default — free, keyless, no quota), openai (needs OPENAI_API_KEY + billing), elevenlabs, or gemini. Match voice names to the model (edge: en-US-AndrewNeural / en-US-AvaNeural; openai: onyx / nova). Fastest path to your first episode — but it rents the voice (except free edge), so for full durability swap to omnivoice.

sesame — Closest open-weight model to NotebookLM voice quality. Requires a CUDA GPU and the [sesame] extras:

uv pip install -e '.[sesame]'

The first run downloads model weights (~6GB). When torch isn't installed, falls back to a silent WAV so the rest of the pipeline still produces a valid episode.

kokoro — Tiny CPU-only ONNX model. Useful when you have no GPU and want something faster than podcastfy:

uv pip install -e '.[kokoro]'

Publishing the feed

Local (recommended for testing)

uv run hexu serve --host 0.0.0.0 --port 8080
# subscribe in your podcast app to http://<your-machine>:8080/feed.xml

S3 / Cloudflare R2 / MinIO

export STORAGE_BACKEND=s3
export S3_BUCKET=my-hexu
export FEED_BASE_URL=https://cdn.example.com/
uv pip install -e '.[s3]'
uv run hexu brief

The publisher writes audio + feed.xml to the bucket, and embeds CDN URLs in the RSS.

Cron setup

# crontab -e
30 6 * * * /Users/you/path/to/hexu/scripts/cron-daily.sh

Logs land under output/logs/YYYYMMDD.log. The wrapper script runs the briefing via uv run so the right venv is picked up.

Cost estimate

Per daily episode, with Gemini 2.5 Pro and Podcastfy/OpenAI TTS:

Component	Cost / episode
Gemini 2.5 Pro (script generation, ~5k input + 1.5k output tokens)	~$0.02
TTS (Podcastfy → OpenAI TTS, ~1000 words, 6-min audio)	~$0.15
Google APIs (Gmail + Calendar reads)	$0 (free tier)
Storage (local)	$0
Storage (S3, ~5MB / episode)	<$0.001
Total	~$0.17 / day, ~$5 / month

Switch TTS_PROVIDER=omnivoice to drive the voice cost to zero with a local OmniVoice backend (free at any volume, voice cloning, 646 languages) — the recommended path. Sesame CSM or Kokoro also drive TTS to zero at the price of GPU electricity or slightly worse voice quality.

Roadmap

• ✅ Voice cloning (via omnivoice — 3-sec reference → personalized hosts).
• ✅ Multi-language (via omnivoice — 646 languages, OMNIVOICE_LANGUAGE).
• Web UI (one-click subscribe, episode browser).
• Slack / Telegram delivery (audio + transcript link instead of RSS).
• Smart deduplication across RSS sources (cluster + summarize related stories).
• Memory of past episodes (don't re-cover the same stories two days running).

Why this exists

When Huxe announced their 2026-05-28 sunset, there was no drop-in OSS replacement. The full background on the decision, the OSS landscape, and the rationale for each component choice is in research/notes/2026-05-22-huxe-sunset-investigation-raw.md (workspace root).

License

This project sits on top of OSS components — see each dependency's license. The hexu glue code itself is provided as-is under the same license as the parent monorepo.