easy-podcast 0.8.1

A local-first, generic-entity podcast knowledge base.

easy-podcast

A local-first, generic-entity podcast knowledge base. Your data lives in a folder you own — durable, human-readable, and portable — with full-text search and a query/graph index on top. Built to grow into transcript search, story segmentation, and cross-podcast connections (same story, author, voice).

Status: v0.7 — feature-complete. The generic store, RSS ingestion, transcription, story segmentation, cross-podcast connections, and a local web UI are all implemented and tested (see What it does). Everything is local and private — no LLMs, no cloud.

Quick start (Docker)

The server ships the phone app inside it. Run it hardened, non-root, exposed only on localhost — one command:

docker compose up -d --build      # serves on http://localhost:8765
# populate your library (via the CLI inside the container):
docker compose exec easypod easy-podcast --data-dir /data add <feed-url>
docker compose exec easypod easy-podcast --data-dir /data sync
docker compose exec easypod easy-podcast --data-dir /data download <podcast_id>

Open http://localhost:8765 and "Add to Home Screen" — that's the installable, offline PWA. To reach it from your phone, the deploy stack (docker-compose.deploy.yml) fronts it with a bundled Tailscale sidecar (HTTPS over the tailnet, no open ports); the full runbook is in docs/SECURE_DEPLOY.md, and a from-scratch, layer-by-layer explainer is in docs/SYNC_DEEP_DIVE.md. Prefer to run it directly (no Docker)? See Install.

Principles — built to respect creators

This is a personal tool for organizing and exploring podcasts you listen to. It is built to always respect human authors and human voice actors:

• No AI training on anyone's content. It honors creators who ask not to be used for AI — their feeds are excluded from examples and tests.
• No LLMs. Transcription (via whisperx) and all analysis run locally and privately; nothing leaves your machine.
• Identify and credit, never imitate. The "connections" features detect and link human authors and voice actors across episodes in order to credit their work — never to clone, synthesize, or impersonate a voice or an author's writing.
• Polite by default. Downloads are rate-limited, use conditional requests, and respect server back-off, so the app never hammers anyone's servers.

Design in one breath

Truth is a Ledger: one JSON file per entity under docs/<kind>/, written atomically. A Lens (an embedded SQLite database, lens.sqlite) is a derived, disposable index over the Ledger providing exact-match queries, full-text search, and referential edges. Delete the Lens and rebuild it from the Ledger at any time. The store is domain-agnostic — any frozen-dataclass Entity is stored and queried through one mechanism.

from easy_podcast import Store, Podcast, Episode, DEFAULT_SPECS

store = Store.open("~/PodcastLibrary")          # the only storage decision; back up = copy the folder
for spec in DEFAULT_SPECS:
    store.register(spec)

show = Podcast.create("https://feeds.example/magnus-archives", title="The Magnus Archives")
store.repo(Podcast).put(show)

store.repo(Episode).put(Episode(
    id=Episode.id_for("guid-42", "https://cdn/ep42.mp3"),
    podcast_id=show.id, title="The Lighthouse", description="A keeper vanishes.",
    audio_url="https://cdn/ep42.mp3",
))

# exact-match query on a promoted field, and full-text search
store.repo(Episode).find(podcast_id=show.id)
for hit in store.search("lighthouse keeper"):
    print(hit.ref, hit.snippet)

Adding a new entity type is a dataclass + one line

from dataclasses import dataclass
from typing import ClassVar
from easy_podcast import Entity, EntitySpec

@dataclass(frozen=True)
class Bookmark(Entity):
    kind: ClassVar[str] = "bookmark"
    episode_id: str = ""
    note: str = ""
    t: float = 0.0

store.register(EntitySpec(Bookmark, promote=["episode_id"], fts=["note"]))
store.repo(Bookmark).put(Bookmark(id="b1", episode_id=ep.id, note="great twist", t=842.0))

No new persistence module, no schema migration.

What it does

The core is pure-Python (feedparser + requests) and the intelligence layers are optional extras you add only if you want them. Nothing uses an LLM; the ML is local transcription and small local embeddings.

Capability	Module	Extra
Generic store (Ledger + Lens), atomic writes, FTS, edges	store/	core
RSS ingestion: fault-tolerant parse, self-throttling verified download, content-based sync, SSRF/decompression-bomb defenses	ingest/	core
Transcription with word-level timestamps (whisperx align) + optional speaker diarization	pipelines/transcribe.py	transcribe
Creator-authored chapters (Podcasting 2.0 podcast:chapters)	pipelines/chapters.py	core
Story segmentation of anthology episodes + author/narrator extraction (heuristics, no LLM)	pipelines/segment.py	transcribe
Local text + voice embeddings → cross-podcast connections (same story / author / transcript / voice)	pipelines/{embed,connect}.py	connections
Full-text + transcript search	store/lens.py (FTS5)	core
Local web UI (localhost-only, CSRF-guarded)	web/	web

Install

pip install -e ".[dev]"          # core: feedparser, requests (pure-python, 3.10–3.13)
pip install -e ".[transcribe]"   # whisperx transcription/diarization (needs Python 3.10–3.12)
pip install -e ".[connections]"  # local embeddings + connection-finding (sentence-transformers)
pip install -e ".[web]"          # local Flask web UI

Transcription needs the ffmpeg binary on PATH. Speaker diarization additionally needs a HuggingFace token (--hf-token or HF_TOKEN).

Command line

easy-podcast add https://rss.acast.com/themagnusarchives   # subscribe
easy-podcast sync                                          # fetch new episodes
easy-podcast download <podcast_id>                         # polite, throttled
easy-podcast transcribe <episode_id> --model base          # + --diarize (needs HF token)
easy-podcast chapters <episode_id>                         # creator chapters, if published
easy-podcast segment <episode_id>                          # split into stories + credits
easy-podcast embed <episode_id>                            # local story embeddings
easy-podcast connections                                   # find cross-episode links
easy-podcast search "lighthouse keeper"                    # full-text + transcript
easy-podcast serve                                         # local web UI on 127.0.0.1:8000

The web UI binds to 127.0.0.1 only — there is intentionally no option to bind a public address, and the app refuses any non-loopback client, so it can't be exposed to a network by accident. The UI is unauthenticated by design; to reach it from another device (e.g. your phone), put a proxy that adds auth + TLS in front of the loopback port — tailscale serve <port> is the easy, private option (only your own enrolled devices can connect).

Listen on your phone

The phone app is a PWA — no app store, no native build. Your library lives on the phone (offline-capable); the home computer keeps the MP3s and does the heavy ML. They sync over an authenticated, encrypted channel, paired once by scanning a QR. The built PWA ships inside the package, so the server serves it with no Node toolchain on the user's side.

1 — set up the server

pip install -e ".[sync]"        # adds the sync server (pynacl + qrcode)
easy-podcast add <feed-url>     # subscribe, then `sync` / `download` as usual

2 — run the sync server. It binds loopback by default — safe, invisible to the network. Reaching the phone is an explicit opt-in; pick a mode:

easy-podcast sync-serve                          # 127.0.0.1 — same machine only (default)
easy-podcast sync-serve --host tailscale \       # off-box: over a private Tailscale net —
    --allow-peer <phone-tailscale-ip>            #   no open port on Wi-Fi, reachable anywhere

Add --tls-cert/--tls-key/--tls-host (a free Tailscale cert) to serve HTTPS — the secure context the phone needs to install it as a real offline PWA. The full step-by-step runbook (Tailscale, a least-privilege jail, certs) is docs/SECURE_DEPLOY.md.

3 — pair the phone. easy-podcast pair "my phone" prints a QR (and a paste-able {"secret":…} line). On the phone, open the server URL and paste that line — or scan the QR with your camera and paste what it decodes — then tap Connect. The app pulls your podcasts/transcripts, fetches audio on demand, plays with lock-screen controls, and "Add to Home Screen" makes it standalone. Pairing a device (or revoking one with unpair) takes effect on a running server with no restart.

Why it's safe (and why an "open port" isn't an exposure)

• Local/overlay clients only, fail-closed: the real TCP peer and the Host header must be loopback / private / your-tailnet — never an arbitrary domain (closing DNS-rebinding). --allow-peer pins it to one exact device, and every refusal is logged.
• Paired devices only. Every request is sealed under the QR secret (authenticated encryption, direction-bound, replay-deduped); the secret is shown as a QR and never crosses the wire. Unpaired → flat 403.
• No discoverable port. Bound to loopback (unreachable) or the Tailscale overlay (the app port lives on a virtual interface nothing physical can address; WireGuard's own socket is silent to non-peers). A scan finds nothing.

A from-scratch, layer-by-layer walkthrough — the protocol, the networking, and a full request trace down to the physical layer — is in docs/SYNC_DEEP_DIVE.md.

Deliberate limits (so it doesn't overclaim): encryption is at the message layer, with no forward secrecy (the pairing secret is long-lived but never crosses the wire); the /audio body leans on the transport (WireGuard/TLS) for confidentiality; a lost unlocked device is out of scope.

Developing the phone app

The PWA is a SolidJS "kernel + one folder per feature" architecture — adding a feature means dropping a self-registering folder under pwa/src/features/, editing no shared file. To build it yourself (a built copy already ships in the package): cd pwa && npm install && npm run build compiles straight into src/easy_podcast/sync/_webapp; npm run e2e drives a real sync end-to-end. Three docs:

• pwa/README.md — start here: how the code works, how to add a feature in one folder, and how to test it (typecheck → Vitest → Node e2e → browser smoke).
• docs/PWA_ARCHITECTURE.md — the deep design ("Ports & SolidJS"): the kernel, the Solid ownership model, every contract, the footguns.
• docs/PWA_FEATURE_CATALOG.md — the ~430-feature backlog with IDs, a dependency map, and the build waves.

Roadmap — delivered

Phase	Scope
0	Generic store (Ledger + Lens), atomic writes, FTS, edges, models	✅
1	RSS ingestion (fault-tolerant parse, rate-limited verified download, content-based sync) + CLI	✅
2	Transcription via easy-whisperx; transcript full-text search	✅
3	Story segmentation (anthology-aware) + author/narrator extraction — no LLMs	✅
4	Text + voice embeddings; cross-podcast connections (same story / author / transcript / voice)	✅
5	Local web app	✅
6	Phone PWA (local-first replica) + authenticated LAN sync — secure by construction	✅
7	PWA feature platform — "Ports & SolidJS" kernel + one-folder-per-feature; features in progress	🚧

License

MIT — see LICENSE.