openpod

<p align="center"> <picture> <source media="(prefers-color-scheme: dark)" srcset="https://raw.githubusercontent.com/openpodhq/openpod/main/assets/brand/wordmark-blue-dark-636.png"> <img src="https://raw.githubusercontent.com/openpodhq/openpod/main/assets/brand/wordmark-blue-636.png" alt="OpenPod" width="212"> </picture> </p> <p align="center"> <img src="https://raw.githubusercontent.com/openpodhq/openpod/main/assets/demo.gif" alt="An AI agent pulls the minutes that matter from hundreds of hours across the shows you follow — every hit timestamped to verify" width="720"> </p> # OpenPod **Pull the ten minutes that matter — to you.** [](https://pypi.org/project/openpod/) [](https://pypi.org/project/openpod/) [](https://github.com/openpodhq/openpod/actions) [](LICENSE) [](#nothing-leaves-your-machine) [](https://modelcontextprotocol.io) Hundreds of hours pile up across the shows you follow. OpenPod pulls out the minutes you're chasing — cited, timestamped to verify, extracted on your own machine — as the starting point your AI agent works from. OpenPod is a local-first agent toolkit — primitives plus packaged skills — that your agent drives over MCP or the CLI. It connects to nothing external; the only corpus is your library on disk. ```bash pip install openpod openpod catch "https://example.com/podcast/feed.xml" ``` **Catch your first episode** — one command, about a minute. You get a timestamped transcript, the key ideas, and jump-to-moment links, all as plain files in a folder you own. `0 servers · 0 accounts · 0 telemetry · $0 to run · test suite runs fully offline` > **Status:** Stage 1 alpha. Local-pure, pull-only. It does what this page shows — nothing more yet. --- ## The problem You subscribe to twelve shows. You're forty episodes behind, every one is three hours long, and the one segment you actually needed last week — you can't find it, because nobody remembers a timestamp. So the backlog grows, the guilt compounds, and "I'll listen on the weekend" becomes a lie you tell yourself on Mondays. Cloud summarizers answer this by asking you to upload what you listen to — and to trust their summary with no way to check it. OpenPod doesn't summarize your listening on someone else's computer. It gives **your** agent the ability to listen, cite, and link — locally. ## What you get Point OpenPod (or your agent) at an RSS feed or a YouTube link. In return: - **A briefing you can verify.** Every claim carries a timestamp and a deep link. Don't trust the summary? Click and hear the source in the native player. OpenPod never re-hosts anyone's audio — it navigates you to the moment. - **A library that compounds.** Everything you catch lands in `.openpod/` as plain files — transcripts, briefings, ideas, notes, clips. Searchable across everything you've ever caught, keyword and semantic, all local. - **Personalized by your own agent.** OpenPod reads your local `persona.md`, so your AI foregrounds the decisions, tools, and numbers *you* care about and drops the rest — the same episode briefs differently for you than for the next person, and no cloud model ever sees who you are. ## Nothing leaves your machine There is no OpenPod server, no account, no telemetry — there is nothing to opt out of, because nothing phones home. The only corpus is your own library on disk. Publisher transcripts and captions are used when they exist; otherwise audio is transcribed locally with Whisper. Your agent brings the intelligence; OpenPod supplies extraction, structure, citations, deep links, and local search. That's the whole trade — and it's why running it costs nothing. ## Your agent can drive all of it OpenPod's toolkit is exposed as an [MCP](https://modelcontextprotocol.io) server, so Claude Code, Cowork, Codex — any MCP client — can catch, search, clip, and brief by name: ```bash openpod-mcp ``` ```json { "mcpServers": { "openpod": { "command": "openpod-mcp", "env": { "OPENPOD_HOME": "/path/to/your/workspace" } } } } ``` Say *"catch me up on this episode"* and the agent runs `catch`, reads your local persona file, and writes a cited briefing back into your library — then hands you the path. Mid-conversation, no dashboard, no copy-paste. ## The four primitives | Tool | Does | Writes | |---|---|---| | `catch <link>` | Ingest → transcribe → structure → brief | `transcript.json`, `ideas.md`, `briefing.md` | | `search <query>` | Retrieve across your local library | ranked cues with deep links | | `export_timestamps <entry>` | Navigable TOC with deep links | JSON / Markdown | | `clip <entry> <start> <end>` | Sentence-snapped local cut + shareable deep-link card | media in `clips/`, card | Every mutating command prints the path it wrote — the output of an action is the address of the artifact it produced. Every read command takes `--json`, shaped identically to the matching MCP tool. ## Skills — the features, by name Primitives are the API; **skills are the product.** Each is a versioned instruction bundle (`openpod skills` lists them; the MCP server exposes them as prompts) that tells an agent how to compose the primitives and report back: | Skill | You ask | It writes | |---|---|---| | **Catch Me Up** | "Brief me on this episode" | `briefing.md`, `ideas.md` | | **Find the Moment** | "Where did they talk about X?" | deep-linked hits | | **What's New** | "What dropped in my feeds?" | digest | | **Cut the Clip** | "Pull the shareable minute" | media + share card | | **Chapter It** | "Give me a navigable TOC" | timestamps (md/json) | | **Follow This** | "Keep me on this show" | `follows.yaml` entry | | **Set Up My Persona** | "Learn who I am" | `persona.md` — yours, local | | **Sharpen My Persona** | "Update what you know about me" | persona Derived block | | **Bring In My World** | "Import my subscriptions" | `imports/`, `follows.yaml` | ## More things it does ```bash # Already have a transcript? Skip the network entirely. openpod catch "https://example.com/ep1" --kind podcast --transcript ./ep1.vtt # Search everything you've ever caught (keyword + local semantic) openpod search "what did they say about raft consensus" # Follow feeds locally; see what's new on your schedule — nothing polls in the background openpod follow "https://example.com/feed.xml" openpod digest # Cut a local, sentence-snapped clip (needs ffmpeg) + a shareable deep-link card openpod clip "test-pod/episode-one-consensus" 320 385 # Import subscriptions from a file you export (OPML) — point-in-time, opt-in openpod import ~/Downloads/overcast.opml # Annotate; notes feed future personalization openpod note "test-pod/episode-one-consensus" "the Raft segment matters for our design" ``` Everything lands in a plain, user-owned tree: ``` .openpod/ persona.md # who you are (evolving, local) follows.yaml # subscribed RSS + YouTube channels library/<show>/<episode>/ transcript.json briefing.md ideas.md notes.md clips/ index/ # local search index (SQLite FTS + embeddings) ``` Want it on two machines? Put the tree in git or Dropbox yourself. OpenPod won't sync it for you. ## Install options ```bash pip install openpod # core: RSS ingest, local search, deep links, clip pip install 'openpod[youtube]' # + YouTube captions & audio pip install 'openpod[asr]' # + local Whisper transcription pip install 'openpod[mcp]' # + the MCP server pip install 'openpod[all]' # everything ``` `ffmpeg` on your `PATH` is required for `clip`. ## Use it as a library ```python from openpod import catch, search result = catch("https://example.com/ep1", transcript_path="ep1.vtt", kind="podcast") print(result.entry_id, len(result.transcript), result.ideas) for hit in search("consensus algorithms"): print(hit.show, hit.start, hit.deeplink) ``` ## Trust & posture FAQ **Does OpenPod download or re-host other people's content?** It navigates, it doesn't republish. Deep links open the moment in YouTube, Spotify, or the open podcast enclosure. Podcast RSS — content published for open distribution — is the primary surface; YouTube support is an optional module that uses official captions, or local transcription of audio you can already access. OpenPod never touches DRM, paywalls, or age gates. **What data do you collect about me?** None. There is no endpoint to send it to. **How is transcription chosen?** Publisher-provided timed transcripts and captions first — free, fast, good for navigation (±2–4s cue timing). Otherwise local Whisper (`asr` extra). Nothing is uploaded either way. `clip` snaps cuts to cue boundaries so clips land on sentences. **Why MIT?** [MIT](LICENSE): maximally permissive and frictionless — install it at work without tripping a license policy, embed it, build on it, no obligations. OpenPod isn't protected by its license; it's protected by being local-pure and simple. No copyleft, no lock-in. **Can I hack on it without heavy models or network?** Yes — the test suite runs fully offline against fixtures. Built by Yoav — building OpenPod in public. ## Development ```bash git clone https://github.com/openpodhq/openpod cd openpod python -m venv .venv && . .venv/bin/activate pip install -e '.[dev]' pytest ``` Contributions that package or connect OpenPod travel furthest: Homebrew formula, Docker image, MCP client recipes, note-app bridges. Open an issue first; response is fast. --- **Long-form is a lot. OpenPod pulls the ten minutes that matter — to you — with a link straight to every one.** ```bash pip install openpod ``` If it saved you an hour this week, [star the repo](../../stargazers) — stars are how the next person finds it.