skimtube

You follow 40 YouTube channels. You watch 4.
skimtube reads the other 36 and leaves you notes.

New videos from the channels you follow become Markdown notes you can skim, grep, drop into Obsidian or hand to your coding agent. No YouTube API key, any LLM (local or hosted), one cron line.

Why

I follow 68 channels, mostly about AI. I was never going to watch them all, and the one video that mattered each week was buried under three thumbnails of someone looking shocked. So now a cron job reads the subtitles for me: in four months it has written 725 notes, and every morning I read one digest instead of a subscriptions page. My coding agent greps the rest.

Side effect: you stop seeing the shocked faces.

What it is good for

• Keeping up with a field without the videos. A daily digest grouped by theme, with links back to each video when something is worth the 20 minutes.
• Giving a coding agent a memory of what is new. Notes are plain files, so grep -ril mcp knowledge-feed/ works, and the Claude Code plugin answers "what did people say about X this month?" with sources.
• Growing an Obsidian vault on autopilot. Point SKIMTUBE_OUT at a vault folder. Every note has YAML front matter (channel, date, URL) for Dataview queries.
• Watching a niche or your competitors. List their channels, add a keyword pre-filter and a topic, and skip everything else. The digest flags when two channels disagree.
• Doing it privately and cheaply. Run it with Ollama for $0, or pay about $0.002 per note with DeepSeek.

What you get

Every few hours it checks the channels you list, grabs the subtitles of new videos, and has an LLM write one structured note per video. Once a day it merges the recent notes into a digest, grouped by theme, with a section for claims that disagree across sources ("channel A says X, channel B says the opposite, same week").

knowledge-feed/
├── INDEX.md                      # one line per video, newest first
├── DIGEST.md                     # latest digest
├── digests/2026-09-21.md
└── 2026-09/
    └── 2026-09-20_some-channel_new-open-model-beats-gpt_abc123.md

Quickstart

pipx install skimtube                      # or: uv tool install skimtube
mkdir kb && cd kb && skimtube init         # writes config.env, channels.txt, blocklist.txt
SKIMTUBE_LLM_API_KEY=sk-or-... SKIMTUBE_LLM_MODEL=deepseek/deepseek-chat skimtube run

init points to OpenRouter by default. For a fully local setup, use Ollama instead. Ollama serves 4,096 tokens of context by default and silently cuts longer prompts, so a normal transcript loses most of its text without any error. Create a model with a larger context first:

printf 'FROM qwen2.5:3b\nPARAMETER num_ctx 16384\n' > Modelfile
ollama create qwen2.5-16k -f Modelfile
SKIMTUBE_LLM_BASE_URL=http://localhost:11434/v1 SKIMTUBE_LLM_MODEL=qwen2.5-16k skimtube run

(Or start the server with OLLAMA_CONTEXT_LENGTH=16384.) On a laptop CPU without a GPU, a 3B model takes about 4 minutes per note; that is fine for a scheduled job, slow for a first try.

The first run only looks back 24 hours (SKIMTUBE_LOOKBACK_HOURS). After that, every run processes whatever is new since the previous one. Then:

skimtube digest                  # merge the last 30 h of notes
skimtube digest --window-hours 168   # weekly

Use it from Claude Code

The repo is also a Claude Code plugin. It adds a skill that finds your feed, searches it, and answers with the video links as sources ("what did the channels I follow say about MCP this month?"). It can also run or set up the collector for you.

/plugin marketplace add arnaugonzalez/skimtube
/plugin install skimtube@skimtube

Other agents (Cursor, Codex, Gemini CLI) can read the same instructions from plugin/skills/youtube-knowledge-feed/SKILL.md.

Why not…

fabric has a much larger prompt library, and tubeless also transcribes videos without subtitles (Whisper) and ranks them by importance. Pick them if you want a daily briefing to read. Pick skimtube if you want the notes to pile up as files that you, Obsidian or a coding agent can grep months later.

How it works

flowchart LR
  A[channels.txt<br/>+ optional subscriptions] --> B[YouTube RSS<br/>per channel]
  B --> C{new video?<br/>keyword filter}
  C --> D[yt-dlp subtitles]
  D --> E[LLM: structured note<br/>or SKIP if off-topic]
  E --> F[knowledge-feed/YYYY-MM/*.md<br/>+ INDEX.md]
  F --> G[digest: themes,<br/>contradictions, weak signals]
  S[(state.json)] <--> C

• No YouTube API key. Channel feeds come from the public RSS endpoint; @handles are resolved to channel ids once and cached in state.json.
• Idempotent. Processed video ids live in state.json, so running it twice does nothing new. If the LLM call fails, the video is retried on the next run (up to 3 times) and the run exits with status 1 so your scheduler shows the failure.
• Notes have a fixed shape: YAML front matter (title, channel, url, video_id, published, collected) and the sections TL;DR, Key topics, What's new, Entities, Facts and figures. Headings stay in English; the content follows SKIMTUBE_SUMMARY_LANG.

Configuration

Values come from config.env in the working directory (or -C DIR), else $XDG_CONFIG_HOME/skimtube/config.env. Environment variables override both.

Your subscriptions (optional). skimtube auth runs a one-time OAuth consent (read-only scope) and from then on merges the channels you are subscribed to. It needs your own Google Cloud OAuth client of type Desktop app (SKIMTUBE_OAUTH_CLIENT_ID / SKIMTUBE_OAUTH_CLIENT_SECRET). While the OAuth app stays in "testing" mode, Google expires the refresh token after a while: re-run auth when the log says so.

Scheduling

run and digest are plain idempotent commands. See examples/ for cron, a systemd user timer, and a GitHub Actions workflow. To publish the feed somewhere else, chain it: skimtube run && rsync -a knowledge-feed/ server:kb/.

To summarize a playlist or a few videos right away, outside the channel feed: skimtube add 'https://www.youtube.com/playlist?list=PL...' -n 30. Videos already in the knowledge base are skipped unless you pass --force.

Limits

• Subtitles are required. Videos without subtitles or auto-captions are skipped.
• Datacenter IPs get bot-checked. YouTube often refuses subtitle downloads from cloud servers ("Sign in to confirm you're not a bot"). A home machine works reliably; on a server you need yt-dlp cookies or a proxy.
• Terms of service. Downloading subtitles with yt-dlp is not a use YouTube explicitly permits. skimtube keeps transcripts in memory only and writes transformed summaries. Don't redistribute transcripts, and check the rules that apply to you.
• Contradictions are leads, not verdicts. Checked by hand over 4 real days, the section flags 2–3 pairs per window and at most one is a genuine disagreement; the rest are nuances the model explains away in its own note.

Non-goals: a web UI, a database, transcribing audio.

Security

Transcripts are text written by strangers, and skimtube hands them to an LLM. If that LLM is an agent CLI with tools (Claude Code, Codex, Gemini CLI…), a video can carry instructions like "ignore the above and run this command", and the agent may do it with your permissions. We reproduced this with claude -p in auto mode. Disable every tool for this job:

SKIMTUBE_LLM_COMMAND=claude -p --tools "" --strict-mcp-config

HTTP providers (SKIMTUBE_LLM_PROVIDER=openai, the default) only return text, so they can't run anything. skimtube itself never runs a shell, opens a port, or sends telemetry; it talks to YouTube, your LLM endpoint and nothing else. See SECURITY.md to report issues.

License

MIT. skimtube is an independent project for YouTube users; it is not affiliated with or endorsed by YouTube or Google.