TeiNam38 downloadsSummarize YouTube captions into structured study notes with Amazon Bedrock, OpenAI-compatible models, Claude or Gemini.
Fetch YouTube captions inside Obsidian, translate and summarize them with the AI model of your choice, and save the result as a study note. It works without a separate summary server, Python, yt-dlp, audio downloads, or speech-to-text.
Local processing here means the plugin itself collects the captions and runs the summary flow. If you pick a cloud model, the captions are sent to that AI provider and usage fees may apply. Point an OpenAI-compatible server such as Ollama or LM Studio at a local address to run the model on your own machine as well.
YY-MM-DD Video title.md file names, upload date, embedded video, and YAML frontmattergit clone https://github.com/TeiNam/obsidian-youtube-llm-summarizer.git
cd obsidian-youtube-llm-summarizer
# Install dependencies
npm ci
# Test and build for production
npm test
npm run build
Copy main.js, manifest.json, and styles.css into .obsidian/plugins/youtube-llm-summarizer/ in your vault, then enable the plugin. If you installed it under the previous ID (obsidian-youtube-summarizer), move that folder's data.json into the new folder to keep your settings. The AWS SDK and the summary prompts are bundled into main.js.
Requires Obsidian 1.13.0 or later. API key authentication works on desktop and mobile. AWS profile and SSO authentication require the Obsidian desktop runtime with Node.js 20 or later. If your runtime is older, update Obsidian with the latest installer.
Click the YouTube icon in the left ribbon, or run Open sidebar from the command palette, to open the sidebar. You can also switch tabs with Enter, Space, and the left and right arrow keys.
Add a YouTube Data API key and channels in settings to see each channel's latest videos, organized by group. Click a video title to open it on YouTube, or click Summarize to summarize it with the progress stage shown on the item. Summarized videos are marked as complete, and you can click the marker to open a note created in the current session. If the channels cannot be loaded, the reason (such as an invalid API key) is shown. Shorts are not listed; only regular videos count toward the per-channel number.
Turn on Auto summarize for a channel under Settings → Monitored channels to have its new videos summarized into notes without clicking anything.
Auto summarize sends captions to your AI provider and uses its paid quota without asking each time. Each daily check also uses two YouTube Data API quota units per channel.
Choose the AI provider and model ID under Settings → YouTube LLM Summarizer. Credentials and model IDs are stored separately for each provider. The API key of the old summary server cannot be used as an AI provider key.
| Provider | Required settings | How it is called |
|---|---|---|
| Amazon Bedrock | AWS region, model or inference profile ID/ARN, and at least one of an AWS profile name or a Bedrock API key | AWS SDK credentials + SigV4 when a profile is set; Converse + Bearer authentication when there is no profile or its credentials cannot be loaded |
| OpenAI / OpenAI-compatible | Base URL, model ID, API key | Chat Completions |
| Anthropic Claude | Model ID, Anthropic API key | Messages |
| Google Gemini | Model ID, Gemini API key | generateContent |
You can type a model ID directly. The default Bedrock example is us-east-1 / us.anthropic.claude-sonnet-4-6; set a region and a model or inference profile that your account can access.
Load models fetches the list for the current provider. For Bedrock it keeps only the latest version of each Claude family (haiku, opus, sonnet) and saves the list in settings. The saved list is used until you load it again, and is shown as a dropdown. Use Enter manually... in the dropdown for a model or inference profile ARN that is not in the list. Listing requires permission to list either inference profiles or foundation models.
Connection test sends a short request with the current settings to check authentication and the model. For Bedrock it tells you whether the AWS profile, the Bearer token, or the Bearer token after a profile failure was used. It consumes a few tokens.
When the plugin falls back to the Bearer token it shows a notice, and it does not retry the profile again within the same summary.
Enter the API key issued in the Bedrock console in Bedrock Bearer token. The Bearer prefix is optional. This field is not for an IAM access key ID or an SSO login token.
For IAM Roles Anywhere, create a profile in ~/.aws/config that sets credential_process to aws_signing_helper, and enter its name. The helper and certificate paths must be absolute paths on this computer.
For SSO, configure a profile with the AWS CLI and sign in.
# Skip the configure step if the profile already exists
aws configure sso --profile my-bedrock
aws sso login --profile my-bedrock
Enter my-bedrock in the plugin's AWS profile field. The SDK reads credentials from ~/.aws/config and the SSO cache. Existing shared AWS profiles also work. When the session expires, renew it with the same login command.
Calls made with a profile are pinned to SigV4, so an AWS_BEARER_TOKEN_BEDROCK environment variable left in your environment cannot take the place of the profile you entered.
Enter the base URL including /v1.
| Server | Example base URL | Model ID / API key |
|---|---|---|
| OpenAI | https://api.openai.com/v1 |
An available model ID / OpenAI API key |
| Ollama | http://localhost:11434/v1 |
The name of an installed model / key optional |
| LM Studio | http://localhost:1234/v1 |
A model ID served by the server / key optional |
| Other compatible servers | The provider's HTTPS base URL | That provider's model ID / API key |
HTTP is allowed only for localhost, 127.0.0.1, and [::1]. Any other address requires HTTPS and an API key. Changing the base URL sends the current OpenAI-compatible key to the new address, so check the address and the key together.
API keys are kept in Obsidian's secret storage, not in the plugin's data.json. You paste a key into its field in settings, the plugin writes it to secret storage under its own name (for example youtube-llm-summarizer-openai-key), and data.json stores only that name. The eye icon only hides the value on screen. Secret storage is local to each device, so enter the keys again on another device. If you use only an AWS profile, the plugin settings store just the profile name, and AWS credentials are read from the existing AWS cache.
Keys that an earlier build saved in data.json as plain text are moved to secret storage the first time the plugin loads, and removed from the file.
~/.aws/config, ~/.aws/credentials, and the SSO cache, and runs the profile's credential_process command (for example aws_signing_helper) if it has one. With the profile left empty, the plugin does not access files outside the vault.Languages that differ only by region, such as ko and ko-KR, are treated as the same, while Simplified and Traditional Chinese are kept distinct. A pasted transcript has no language information, so it goes through the translation step.
The translation and summary prompts live in src/prompts/ and are bundled into main.js. They detect the genre, follow study-note instructions, and produce six sections such as ===GENRE=== that the plugin parses into the note. Summaries in languages other than Korean keep the same structure with an added output-language instruction. The note's fixed section headings follow the summary language for Korean, Japanese, Simplified Chinese, and Traditional Chinese, and are English for every other language.
When automatic captions appear varies by video, and there is no guarantee they are generated within a day. If captions are missing or incomplete, the plugin does not fall back to speech recognition; it tells you to retry or paste a transcript. You can also use Script / transcript when YouTube throttles requests or requires login or age verification. If the video information cannot be fetched in that case, the video ID is used as the title.
Caption collection depends on YouTube's public player response, so it can be affected by changes on YouTube's side or by request limits.
| Setting | Description |
|---|---|
| Maximum input characters | Default 200,000. If the captions or translation exceed it, the run stops; text is never silently cut |
| Maximum output tokens per call | Default 65,536. Adjust within the output limit your model supports |
| Summary language | Default Korean. Set independently of the UI language. Sets the language of the note body and, for Korean, Japanese, and Chinese, its section headings (English otherwise) |
| YouTube Data API key | For adding channels and the latest-video feed. Not needed to summarize a URL directly |
| Save folder / Subscription save folder | Where regular and feed summaries are saved |
| Channel save folder / Group | Per-channel save location and feed group |
| Videos per channel | The latest 1–10 videos, not counting Shorts |
| Auto summarize (per channel) | Off by default. Summarizes the channel's new videos one day after upload; checked once a day |
The input character count is not a token count, so set it to fit your model's context window. If a translation or summary response is cut off by the output limit, no note is saved. Each model request waits up to 3 minutes; Obsidian's requestUrl has no network cancellation API, so the provider may still finish processing after a timeout.
Existing channels, folders, groups, and summarized-video records are kept. The old server key is also preserved in case you go back to an earlier version, but it is never sent on the new summary path. After updating, set the AI provider's credentials and model.
npm run dev # Watch build
npm run lint # Obsidian plugin guidelines (eslint-plugin-obsidianmd)
npm test # Caption, summary, authentication, note, and feed tests
npm run build # Type check + release bundle
Pushing a tag with the same name as version in manifest.json (for example 1.0.0) makes GitHub Actions run the lint and tests, build, generate a build provenance attestation, and attach main.js, manifest.json, and styles.css to a draft release.
The core implementation lives in YouTubeCaptionClient.ts (captions), AiModelClient.ts (authentication and model calls), SummaryEngine.ts (translation and summary format), and SummarizerService.ts (the whole flow through saving the note).