MinusPod

MinusPod serves re-cut RSS feeds per podcast, so any app that supports a custom RSS URL works — Apple Podcasts, Pocket Casts, Overcast, Spotify, and the rest just point at the MinusPod feed. It also emits Podcasting 2.0 transcripts, chapters, value-for-value tags, and AI-content disclosure, and uses Podping so your app refreshes within seconds of a new episode, not on a polling schedule.

Yes. The default docker-compose assumes an NVIDIA GPU for local Whisper, but MinusPod ships a multi-arch CPU image (docker-compose.cpu.yml) that runs natively on amd64 and arm64. You can also offload Whisper entirely to a remote API — Groq, OpenAI Whisper, whisper.cpp, or OpenVINO on an Intel GPU — so a small home server with no discrete GPU is enough.

Both work. You can point MinusPod at Anthropic, OpenRouter, or any OpenAI-compatible endpoint, or run fully local inference with Ollama. Per-stage settings (model, temperature, token limits, prompts) are editable, and the LLM Benchmark Report in the repo lists per-model F1, latency, and cost so you can pick a tradeoff that matches your hardware and budget.

When you correct or confirm an ad detection in the web editor, MinusPod extracts a pattern from that span and promotes it up the scope chain: podcast to network to global. Repeat sponsors then get caught automatically on new episodes without re-querying the LLM. There is also an opt-in community pattern sync, with changes submitted back via PR, so popular sponsors across many feeds can be blocked from day one.

Provider keys are encrypted with AES-256-GCM behind a MINUSPOD_MASTER_PASSPHRASE that you set on first boot — without it, keys sit in the database as plaintext. Scheduled backups are also encrypted and tied to the same passphrase, so restoring requires the original secret. The web UI is gated by a single shared password with login lockout, and the app is designed to run behind a reverse proxy or a Cloudflare tunnel for remote access.