NotebookLM Research
Use the installed commands or the 13 MCP tools to turn user-provided sources into
grounded answers and NotebookLM artifacts. Commands emit JSON on stdout and progress
or diagnostics on stderr, so preserve stdout when another tool will consume it.
This integration uses NotebookLM's browser session and unofficial web API through
notebooklm-py. Do not promise that Google-side availability, quotas, or generation
time are stable.
Authentication
Prefer the profile-aware helper:
notebooklm-auth setup
notebooklm-auth verify
Use notebooklm-auth setup --browser chrome --fresh when the user explicitly
wants the locally installed Google Chrome instead of bundled Chromium.
For a zero-install login:
uvx --from notebooklm-py notebooklm login
Profiles are supported through --profile NAME before the subcommand or through
NOTEBOOKLM_PROFILE. Current sessions are normally stored below
~/.notebooklm/profiles/<profile>/storage_state.json; never read, print, copy, or
commit that file. If authentication expires, run setup again.
Core CLI
Create a notebook from mixed sources:
notebooklm-skill create \
--title "AI safety evidence" \
--sources https://example.com/article https://youtu.be/example \
--files ./paper.pdf \
--text-sources "A user-supplied observation" \
--strict
Inspect and ask:
notebooklm-skill list
notebooklm-skill list-sources --notebook "AI safety evidence"
notebooklm-skill summarize --notebook "AI safety evidence"
notebooklm-skill ask --notebook "AI safety evidence" --query "What findings conflict?"
Add exactly one source:
notebooklm-skill add-source --notebook "AI safety evidence" --url https://example.com/new
notebooklm-skill add-source --notebook "AI safety evidence" --file ./appendix.docx
notebooklm-skill add-source --notebook "AI safety evidence" \
--text "Raw notes" --text-title "Interview notes"
Run NotebookLM web research and import results:
notebooklm-skill research \
--notebook "AI safety evidence" \
--query "Recent empirical evaluations" \
--mode deep --max-sources 10
Use --no-wait for a task ID without waiting. Use --no-import-results when the
research results should not become notebook sources.
Notebook titles may be used only when they resolve uniquely. Prefer IDs in automation.
Artifact generation
Supported canonical types:
audio, video, cinematic, slides, report, study-guide, quiz,
flashcards, mind-map, infographic, data-table.
Generate and optionally download in one operation:
notebooklm-skill generate \
--notebook "AI safety evidence" \
--type slides --lang zh-TW \
--slide-format presenter-slides \
--output ./output/deck.pptx --output-format pptx
Long media jobs can be detached and downloaded later by exact ID:
notebooklm-skill generate --notebook NOTEBOOK_ID --type audio --no-wait
notebooklm-skill list-artifacts --notebook NOTEBOOK_ID --type audio
notebooklm-skill download --notebook NOTEBOOK_ID --type audio \
--artifact-id ARTIFACT_ID --output ./output/podcast.m4a
Convenience commands:
notebooklm-skill podcast --notebook NOTEBOOK_ID --output podcast.m4a
notebooklm-skill qa --notebook NOTEBOOK_ID --difficulty hard --output quiz.json
Generation supports per-type options. Inspect the live contract before composing an
unfamiliar call:
notebooklm-skill generate --help
Existing output files and symlinks are rejected. Use --force only when the user
explicitly wants an overwrite. Quiz and flashcard downloads support JSON, Markdown,
or HTML; slide downloads support PDF or PPTX.
High-level pipelines
notebooklm-pipeline research-to-article \
--sources https://example.com/a https://example.com/b \
--title "Evidence review" --language zh-TW --audience "engineers"
notebooklm-pipeline research-to-social \
--sources https://example.com/a --platform threads --variants 3
notebooklm-pipeline batch-digest \
--rss https://example.com/feed.xml --max-entries 20 --qa-count 5
notebooklm-pipeline generate-all \
--files ./paper.pdf --types audio slides report mind-map \
--output-dir ./output --artifact-concurrency 2
trend-to-content requires a separately installed trend-pulse command. Override
its executable safely with TREND_PULSE_CMD; the integration does not invoke a shell.
Pipelines create drafts and local artifacts. They do not publish to social networks,
CMS products, or other remote destinations.
MCP server
Start stdio mode for an MCP client:
notebooklm-mcp
Example configuration:
{
"mcpServers": {
"notebooklm": {
"command": "uvx",
"args": ["--from", "notebooklm-skill", "notebooklm-mcp"]
}
}
}
Available tools:
nlm_create_notebook, nlm_list, nlm_delete
nlm_add_source, nlm_list_sources
nlm_ask, nlm_summarize
nlm_generate, nlm_download, nlm_list_artifacts
nlm_research, nlm_research_pipeline, nlm_trend_research
Notebook deletion requires confirm=true. HTTP mode binds only to loopback:
notebooklm-mcp --http --host 127.0.0.1 --port 8765
Do not expose HTTP mode directly to a network. If remote access is unavoidable, put
it behind an authenticated TLS proxy and apply host-level access controls.
Operating rules
- Verify authentication before a long workflow.
- Confirm sources were ingested; treat
partial or failed source entries honestly.
- Ask focused questions and retain returned citation metadata.
- Use exact notebook and artifact IDs in repeated automation.
- Use bounded source/artifact concurrency; generation is quota-sensitive.
- Do not delete notebooks or overwrite output without explicit user intent.
- Do not claim a draft was published; no publishing integration exists here.
- Return the JSON result or a faithful summary, including partial failures.
Exit codes and recovery
0: operation completed successfully.
2: invalid or ambiguous arguments.
4: authentication required.
1: upstream, network, generation, or other operational failure.
130: interrupted by the user.
Common recovery:
notebooklm-auth verify
notebooklm-auth setup # missing or expired session
notebooklm-skill list-artifacts --notebook NOTEBOOK_ID # inspect a timed-out job
Use notebooklm-skill --help, notebooklm-pipeline --help, and the relevant
subcommand's --help as the authoritative local command contract.