Let your AI agent quote what was said in a podcast or video, with the minute

An MCP server: works with Claude Code, Cursor, and Codex.

Give your agent a podcast or video link, or just a topic, and it gets the words of the part that matters, with the minute each was said, turned into text on your own computer.

Try it

ASK YOUR AGENT

Find a podcast episode on speculative decoding and quote what it says about the draft model, with the timestamp.

Behind the scenes

omniseek_search(query="speculative decoding", sources=["apple_podcasts"])
# picks an episode from the returned feed, then:
omniseek_transcribe(url="https://www.buzzsprout.com/2529712/episodes/18794851-speculative-speculative-decoding.mp3",
                    start="1:00", duration="0:45", language="en")

WHAT CAME BACK

1:00 to 1:45“those older systems used a fast, small draft model to guess a few words ahead and then a massive target model came in”

From a real run on 2026-10-10. The first time, it downloads a speech-to-text model (about 900 MB). After that, transcripts are saved and reused.

Tip: ask for the part you need, say minutes 40 to 50, not the whole two-hour episode.

Install

TWO COMMANDS

uv tool install "omniseek[asr]" --with torch --with torchaudio
claude mcp add omniseek -- omniseek

Turning speech into text needs this optional add-on (the asr extra and PyTorch). Details and disk sizes: Podcast and video transcription.

Linux

Without an NVIDIA GPU, add --torch-backend cpu to the first command, or PyTorch downloads several GB of CUDA libraries you will not use.

Where it looks

apple_podcasts: find shows and episode feeds by topic
xiaoyuzhou: Chinese podcast episodes
chinese_podcasts: Chinese tech and research podcast feeds
youtube: videos, transcripts, and top comments
bilibili: Chinese videos
podcast_index: podcast directory (needs a free key)

Every source, with whether it needs a key or a login: docs/sources.md. Tool reference: docs/tools.md.

OmniSeek · GitHub · PyPI · Apache-2.0