@michaelkory/notebooklm

A bun CLI (zero runtime deps) that builds and queries Google NotebookLM notebooks programmatically. It drives NotebookLM over raw Chrome DevTools Protocol against a Chrome you're already signed into — no API key, no extension, no reimplemented auth. The in-page calls inherit your existing session.

Build notebooks from a YouTube channel, playlist, or search; from a person's teaching videos + interviews; from a cross-creator topic; from blogs; or from LinkedIn articles. Discover the directory channels and experts behind a niche, subscribe to channels, mine your watch history, and ask grounded, citation-backed questions.

⚠️ Unofficial. This automates NotebookLM's private web backend from your own logged-in session. Keep volume human-scale; everything depends on the one signed-in tab. Not affiliated with Google. (LinkedIn/YouTube scraping is likewise ToS-sensitive — go gently.)

Install

bunx @michaelkory/notebooklm <command>      # no install
bun add -g @michaelkory/notebooklm          # or: global `notebooklm` binary

Prerequisites

• bun installed (runs the TypeScript directly).
• A logged-in NotebookLM Chrome reachable over CDP — default http://127.0.0.1:9224 (override with NBLM_CDP). For LinkedIn/YouTube features that Chrome should also be signed into those.
• list is the health check — if it errors with "no CSRF token" or a connection failure, open/sign into NotebookLM in that Chrome.

Commands

Harvest / inspect (read-only — return data, create nothing)

| Command | Action | |---|---| | list | Notebooks [{id,title,emoji,source_count}] — also the health check. | | harvest <channel> [maxScrolls] | A channel's videos (handle/URL/id) + duration/views/age. | | harvest-playlist <playlistUrl> [maxScrolls] | A playlist's videos, in order. | | harvest-search <query…> | YouTube search results (title/duration/views/channel). | | harvest-person <name> [ownChannelHandle] | A person's teaching videos (own channel + interviews, filtered). | | harvest-blog <rootUrl> [limit] | Article URLs from a site (sitemap → RSS fallback). | | harvest-linkedin <profileUrl> | A LinkedIn profile's articles → text cached as markdown. | | history <takeout.zip\|watch-history.json> [sinceISO] | Your watch history → channels ranked by views (music filtered). | | caption-check <videoUrl…> | Whether each video has captions. |

Build notebooks (always create new; record a manifest)

| Command | Action | |---|---| | build <label> <channel> [maxScrolls] [filters] | From a channel. | | build-urls <label> <url…> | From explicit URLs. | | build-directory <label> <channelNameOrHandle> [maxScrolls] | Resolve real channel → harvest → drop Shorts → 300 most-recent. | | build-person <label> <name> [ownChannelHandle] | A person's teaching corpus. | | build-topic <label> '<q1\|q2\|…>' | Cross-creator topic (teaching-filtered, \|-separated queries). | | build-linkedin <label> <profileUrl> | LinkedIn articles → pasted-text sources. |

build filters: --limit N --min-duration SEC --max-age-months M --min-views N --include kw,kw --exclude kw,kw

Discover / curate

| Command | Action | |---|---| | find-directories <seed…> | Rank channels by how many of your seed people they host (interview directories). | | person-hosts <name> [ownChannelHandle] | The channels that host a given person. | | find-experts <topic> <notebookId…> | Ask each notebook who the topic's experts are → per-notebook answers + citations. | | subscribe <name\|handle…> | Subscribe to a channel (resolver-guarded). Unsubscribe must be done manually — YouTube needs a real click. | | delete-notebook <id…> | PERMANENTLY delete notebooks (no undo). | | ask <notebookId> <question…> | Grounded answer → {answer, citations}. |

bunx @michaelkory/notebooklm build-directory "My First Million" "My First Million"
bunx @michaelkory/notebooklm find-directories "Alex Hormozi" "Rob Walling" "Dan Koe"
bunx @michaelkory/notebooklm find-experts "first 100 SaaS customers" <nbId1> <nbId2> …
bunx @michaelkory/notebooklm build-linkedin "Justin Welsh" https://www.linkedin.com/in/justinwelsh/

Configuration

| Env var | Default | Purpose | |---|---|---| | NBLM_CDP | http://127.0.0.1:9224 | CDP endpoint of your logged-in Chrome. | | NBLM_NOTEBOOKS_DIR | ~/projects/ideacode/Atlas/NotebookLM | Manifests, per-notebook markdown notes, and the LinkedIn text cache (sources/linkedin/). |

How it picks sources (the curation model)

• Teaching density over views. Ranking favors instructional content (how-to / framework / playbook / teardown), not popularity; views are at most a tiebreak, never a gate.
• Own-channel-first for teachers; search for the channel-less. A teacher's lessons live on their channel; people like Seth Godin / Justin Welsh live in others' interviews + their writing — so person/topic notebooks pull from search and LinkedIn too.
• Directories = multi-guest interview shows, found by who repeatedly hosts your seeds.
• Channel resolution is squatter-guarded — resolveChannel requires ≥5k subs + a distinctive-name match and picks the highest-sub result, so a 3-video imposter can't slip in.
• NotebookLM cap: 300 sources / notebook. build-directory takes the 300 most-recent long-form videos. NotebookLM does not reject captionless videos (verified).

Notes

• Always creates NEW notebooks — no get-or-create matching.
• Builds record a canonical manifest.json (keyed by notebook id) + a per-notebook markdown note — handy for an Obsidian vault.
• LinkedIn URLs can't be ingested by NotebookLM directly (auth wall), so build-linkedin scrapes the text via your session, caches it as markdown, and adds it as pasted text.
• Runtime: bun only (raw CDP over bun's native WebSocket; no Playwright, no Node).
• Maintainer release steps: see PUBLISHING.md.