kura

Turn YouTube into notes you'll actually re-read.

Paste a URL. Get a structured Markdown file — cliff notes, quotes with timestamps, and further reading — dropped straight into your Obsidian vault or Notion database.

Get started

Writes to your vault, not another silo

A .md file with YAML frontmatter. Drop it in Obsidian, Notion, Logseq, or any folder of Markdown. No lock-in. Your notes live in your filesystem, indexed by the tool you already trust.

Timestamped quotes link back to YouTube

Every quoted passage in the summary includes a clickable timestamp. One click takes you back to the exact moment in the video — verify a claim, re-listen, or dig deeper.

Open-source, bring your own key

The entire app is on GitHub under MIT. Fork it, self-host it, swap the model, change the templates. Plug in a Gemini or Claude API key and run it on your own hardware.

From fifteen minutes to thirty seconds

Manual workflow
  1. 01Open YouTube, enable captions.
  2. 02Copy the transcript from the CC panel.
  3. 03Paste into ChatGPT, write a prompt.
  4. 04Review, fix errors, trim irrelevant sections.
  5. 05Copy output into your vault.
  6. 06Add frontmatter, tags, source URL by hand.
  7. 07Link back to video timestamps manually.
~15 min
With Kura
  1. 01Paste URL.
  2. 02Pick a template.
  3. 03Download or sync.
~30 sec

Open-source core

GitHub starsLicense

The transcript pipeline, Markdown renderer, Obsidian plugin, and backend are on GitHub under MIT. Run the whole thing on your own hardware with your own Gemini or Claude API key — no quotas, no billing, no lock-in.

View on GitHub →Self-host guide →

Frequently asked

Does this work with my existing Obsidian vault structure?
Yes. Kura produces a standard .md file with YAML frontmatter — the same format Obsidian uses natively. Drop it in any folder, use any existing tag taxonomy, keep your current templater setup. Kura does not assume anything about your vault layout.
Can I self-host? What about yt-dlp?
Yes. The core pipeline is on GitHub. For self-hosting, you bring your own Gemini or Claude API key. The open-source version includes a yt-dlp fallback for transcript extraction when the hosted transcript provider is unavailable — useful if you care about resilience or run into region-blocked videos.
Does re-processing the same URL create a duplicate?
No. Kura dedupes server-side per user — submitting the same URL twice returns the existing summary instead of re-running the pipeline.
Which models does Kura use?
The default is Gemini 2.5 Flash for structured output. You can swap in any model that supports structured JSON output — Claude, GPT-4o, local models via an OpenAI-compatible endpoint. Model choice is a config flag.
Is my data private?
Transcripts and summaries are stored in your account's database row. They are not used to train any model. When you self-host, nothing leaves your machine except the call to the LLM provider you chose.