Skip to main content

Release history and per-version changelog

All notable changes to this project are documented in this file.

The format is based on Keep a Changelog, and this project adheres to Semantic Versioning.

[3.2.0] - 2026-09-03​

Added — content_delete: generated content can finally be removed​

DELETE_STUDIO had sat in the RPC id table since v3, declared and called by nothing. Content could be generated and listed but never deleted, so a notebook accumulated every draft anyone ever asked for and the only way out was the web UI. This is the second piece of dead configuration this area has turned up, and the second one found by needing it rather than by reading.

content_delete (DELETE /content/:contentId) removes an audio overview, video, report, infographic, presentation, data table, flashcard set or quiz. RPC-backed, no browser, single id per call — batch shapes were probed against the live server and refused. Verified by deleting nine real artifacts and confirming the notebook came back empty with its source untouched.

Note-backed mind maps are deliberately out of scope: they live in the notes system rather than the Studio library, and deleting one only clears it. Payload and that carve-out from teng-lin/notebooklm-py (MIT).

Closed — both open issues, settled by measurement rather than by waiting​

#34 (generated content in the wrong language) is verified end to end on the path the reporter was actually running: an audio overview requested in Spanish, on the browser transport, from a notebook with no Spanish sources, came back as "Física del magma y formas volcánicas". That was the last unproven case, and it needed 3.1.2 — 3.1.1's fix handled the report-shaped dialog and missed the audio one, which was the filed case.

#30 (idle server accumulating CPU) is closed on a measurement instead of an unfalsifiable hypothesis. A bare launch with stdin held open and silent, on 3.1.2:

elapsedcumulative CPU
130 s1 s
592 s1 s
671 s1 s

That one second is startup and it does not grow — against the 360 minutes over six hours in the report. Every libuv worker idle, no browser spawned alongside, clean shutdown on SIGTERM.

[3.1.2] - 2026-08-21​

Fixed — 3.1.1's browser-path language fix covered one dialog shape out of two​

It was written against the report dialog, where the language menu hides behind each format tile's pencil, and it looked for that pencil first. Audio Overview and Video Overview have no pencil — their card opens straight onto a customisation panel with the language menu already on it. So the very case the original report was about, an audio overview in the wrong language, was still not fixed.

The menu is now used when it is already present, and the pencil is only sought when it is not. Verified live on the audio dialog: 80 options listed, the control moved from English to español.

Two more faults found by testing it rather than trusting it:

  • The wait was far too short. These dialogs have been measured taking more than ten seconds to render, and a three-second check does not report "no language menu" — it invents one, and then generates in the wrong language while claiming to know why.
  • The generate button was searched page-wide, so it matched a stray button behind the overlay and the panel was declared to have none. The language had in fact been set, so the caller got a warning about a failure that had not happened — the same false-report defect this whole release is about, produced by the fix for it.

Verified end to end across both dialog shapes and four languages: data table in Italian, report in Japanese and Brazilian Portuguese, audio menu selection in Spanish.

Added — an unapplied language is now said out loud​

When a language is requested and NotebookLM's browser UI offers no way to set it for that content type, the result carries a warnings entry saying the content came back in the account's default language and that the RPC transport can set it. Content still generated, caller still told.

[3.1.1] - 2026-08-21​

Fixed — generated content came out in the server's interface language, not the one you asked for​

Reported in #34 by @Javiergomezdev: a Spanish notebook, a Spanish account and language: "Spanish" produced a French audio overview and a French mind map, while the tool reported status: "ready". Nothing signalled the artifact was unusable; it surfaced only when someone opened the file.

Three separate faults were stacked here.

The interface locale was choosing the content language. uiLocale exists so the browser fallback's text selectors match the page they are reading. It has no business deciding what language an artifact is written in, yet content_generate fell back to it, and generate_study_aid used it outright with no way to override. Output language is now its own setting — NOTEBOOKLM_CONTENT_LANGUAGE, defaulting to English — and every generator takes a per-call language.

The language argument was documented in a form NotebookLM never accepts. The tool told callers to pass "Spanish"; NotebookLM identifies an output language by a BCP-47 code. A name landed in the payload slot, matched nothing, and the request's hl silently decided instead. All 81 supported languages are now resolved from a code (es, pt_BR, zh_Hans), an English name ("Spanish"), or a native name ("Español"), and anything unrecognised is refused rather than quietly generated in another language. The browser path gets the native name, which is what NotebookLM's language menu actually lists — an English name never matched an option there either.

hl on the request outranks the payload. Verified live: a mind map carrying "es" in exactly the slot the protocol specifies came back in German, because the request went out as hl=de. Generation calls now send the content language as hl; the same request that returned German now returns Spanish. The mind-map payload was also sending an empty language slot and an empty prompt slot, both of which are now filled — generate_mind_map gains language and focus.

NOTEBOOKLM_UI_LOCALE no longer discards an unsupported value in silence. Setting it to es left you on fr without a word, which is what made this take a long debugging session to pin down. It now says so, and says where the content-language setting lives.

The default interface locale is now English, not French. The old default was justified in a comment as "the most common Google Account language", which was never true. If you were relying on French output, set NOTEBOOKLM_CONTENT_LANGUAGE=fr.

Fixed — the browser fallback ignored language too (#36)​

On the browser path the language argument never reached NotebookLM: the format dialog has no language control, and the code never went deeper. It sits one level down, behind each format tile's pencil button, next to a free-text description field. Generation now goes through that panel whenever a language is asked for, and falls back to the plain preset click if the panel is not offered.

Verified on the same call, same notebook, same transport:

before → "Analysis of Volcanic Formation and Structural Classifications"
after  → "Análisis de la Formación y Tipología de los Volcanes"

Nine names in the language catalog were corrected against the menu as it actually renders — it says Indonesia where the reference says Bahasa Indonesia, Latin where it says Latina. A name that does not match the option text is a language silently not applied, which is the bug this release exists to remove. Options are matched on their own text rather than by substring, so Español can no longer select Español (Latinoamérica).

Fixed — a citation with no excerpt claimed the filename as its quotation​

When the excerpt behind a citation marker could not be extracted, the source name was put in the excerpt field. Callers could not tell a real quotation from a stand-in, and the formatters printed the filename inside quotation marks — [1: "Paper.pdf"] reads as though the source says "Paper.pdf".

An absent excerpt is now absent: sourceText is empty, and each format says so in its own way ([1: Paper.pdf — no excerpt], (no excerpt extracted)). The vault writer already handled the empty case correctly, so its output becomes accurate with no schema change.

Fixed — manage_labels was calling the wrong RPC and never worked​

Every label operation returned RPC "LABEL_MANAGE" returned no result — the id likely rotated, listing included. The id had not rotated: listing labels is its own RPC (I3xc3c / GetLabels) and we were using the create id (agX4Bc) for both, then reading the result at the wrong index — create echoes the label list one slot further along than list does. Found while investigating #34, fixed and verified live: create, list and delete now round-trip.

Fixed — "the id likely rotated" was reported for calls the server had simply refused​

content_generate, generate_study_aid and generate_mind_map were failing with RPC "CREATE_STUDIO" returned no result — the id likely rotated, pointing at NOTEBOOKLM_RPC_OVERRIDES and a replacement id to go and find.

There was no rotated id, and nothing upstream was broken. NotebookLM was answering normally and refusing the call — PERMISSION_DENIED, because the notebook was shared with the account rather than owned by it, and NotebookLM only lets you generate content in your own notebooks. The refusal travels in the wrb.fr envelope at index 5 as a gRPC status code, and the client was throwing that away: it only looked for a string payload, found none, and concluded the id was unknown.

A refusal and a rotated id are now different errors, because they carry different evidence — a refusal comes with an envelope for the id, a rotated id comes with none at all. Refusals name the status and what it means for you (PERMISSION_DENIED, RESOURCE_EXHAUSTED when a daily quota is spent, NOT_FOUND, UNAUTHENTICATED, …), and are no longer retried, since a verdict does not change in a second — except UNAUTHENTICATED, which is exactly the stale-token case the retry was built for.

The same message had already caused one wrong diagnosis, and the label bug above was found under it too. Status-code semantics confirmed against teng-lin/notebooklm-py's decoder (MIT).

[3.1.0] - 2026-08-20​

Added — read what NotebookLM actually indexed (source_list, source_read)​

Two new tools, both RPC-backed (no browser):

  • source_list (GET /notebooks/:id/sources) — every source in a notebook with its ID and title. Until now nothing exposed source IDs, so nothing could address an individual source; a 153-source notebook answers in about a second.
  • source_read (GET /notebooks/:id/sources/:sourceId) — a source's full indexed content: the exact text NotebookLM reasons over, which the web UI only ever shows in fragments. Quote a source verbatim, verify what a PDF or a web page really yielded, or pass the raw material to another tool.

source_read takes source_id or source_name (partial, case-insensitive). A name matching several sources is an error naming the first few, never a guess — silently reading the wrong document is worse than being asked again.

format: text (default) returns the plain-text rendition; format: html keeps headings, lists and links but inlines images as base64, so it is markedly bigger.

Long sources are paginated. The first PDF measured here ran to 157,716 characters — enough to spend a caller's whole context before it could judge whether it wanted the document. So source_read returns 20,000 characters at a time by default, alongside totalChars, nextCursor and a continue instruction naming the exact next call. Pages are cut at the nearest line break (the nearest tag close, in HTML) so none ends mid-word, and concatenating every page reproduces the document byte for byte — verified live across nine pages of that 157,716-character source. paginate: false returns the whole thing in one response, and max_chars sets the page size.

The GET_SOURCE render selectors and envelope positions were read from teng-lin/notebooklm-py (MIT), which documents this endpoint; the implementation is our own. Verified live against a 153-source notebook.

Fixed — GET / reported version 1.5.2 and linked to someone else's repo​

The server descriptor read process.env.npm_package_version, which is only set when the server starts through an npm script; started as node dist/http-wrapper.js — the documented way, and what the Docker image does — it fell back to a hardcoded 1.5.2. It now reads the real version, and its docs link points at this project's documentation instead of the upstream fork it was copied from.

Fixed — the OpenAPI spec was missing 12 endpoints and stuck at 1.5.9​

deployment/docs/openapi.yaml documented 31 of the 44 REST endpoints: every v3 extension (sharing, labels, study aids, mind maps, research), the notes read/list routes, and /batch-to-vault were absent, and info.version had never moved past 1.5.9. Spec and code now match endpoint for endpoint, and version:sync keeps info.version in step so CI blocks the drift from returning.

[3.0.4] - 2026-08-19​

Fixed — the RPC transport was silently dead since the rebrand​

Every RPC call reported Not authenticated — session expired on a perfectly valid session and fell back to the DOM scrape, so the headline feature of 3.0.0 — driving NotebookLM's internal batchexecute API, 10-100x faster than scraping — has been inert. The dual-transport design masked it completely: answers stayed correct, only slower, so nothing looked broken from outside.

Cause: the July 2026 rebrand moved accounts to notebook.google.com, but the client collected its cookie jar for notebooklm.google.com and talked to that host, which answers a logged-out homepage for such an account. The jar and the target host have to match.

Fixed by trying notebook.google.com first and falling back to the legacy host for tenants still served there, keeping whichever bootstraps. Verified live on an authenticated account: notebook_list logged RPC list failed … falling back to DOM scrape before, and RPC transport bootstrapped with no fallback after.

Anyone on 3.0.0-3.0.3 has been running the browser path for every operation. This restores the advertised performance with no configuration change.

[3.0.3] - 2026-08-19​

Fixed — wrong answer returned in long conversations (virtualized chat list)​

notebook_ask with a session_id could return an older, unrelated answer instead of the one just generated, on any notebook with enough chat history. Diagnosed by @Astraktes (#29): NotebookLM's chat list is virtualized, so the mounted DOM window does not keep order in sync with message recency — a freshly generated answer was observed at index 8 of 11, with older, already-known answers at 9 and 10. Every position rule (last container, or index past a pre-submission baseline count) therefore picks the wrong container.

Reverting to plain hash dedup was not an option: that is what 2.2.0 replaced, because a repeated answer text (two questions both answered "Yes") collides with a known hash and times out after the full 5-minute budget.

Both are fixed by counting occurrences instead of testing membership. The baseline is now a multiset (snapshotAnswerTexts → text hash to occurrences), and a mounted answer is new when its text occurs more often than the baseline recorded. A second "Yes" goes 1 to 2 and is detected; an unchanged older answer does not move and is not. DOM position is never consulted.

  • countAnswerContainers and baselineContainerCount removed outright, call sites in browser-session and content-generator migrated in the same change.
  • Both failure modes are now regression-tested.
  • Known residual: a message unmounted when the baseline was captured is absent from it and reads as new when it scrolls into view. This now logs a warning naming the incomplete baseline instead of failing silently.

Only the browser path was affected — one-shot notebook_ask calls go over RPC and were always correct.

[3.0.2] - 2026-08-19​

Fixed — --help started a server instead of printing help (#30)​

notebooklm-mcp --help booted a full stdio MCP server and stayed alive until killed, because the subcommand dispatch added in 3.0.1 matched bare words only (!subcommand.startsWith('-')), so every flag fell past the dispatch table into the server. help worked; --help could not. Launched from a terminal, stdin never reaches EOF, so nothing shut the stray server down.

  • --help / -h and --version / -v are handled before the subcommand table and exit; an unknown subcommand still exits 1.
  • An unrecognised flag no longer passes silently — it is reported on stderr and the server still starts, since MCP clients sometimes append flags of their own and aborting would break them. Warnings stay on stderr; stdout is the JSON-RPC channel.
  • Help text now documents the installed-binary commands, not only the npm run forms that apply to a cloned repo.

Fixed — Docker image failed to start when built from a Windows clone​

The image runs scripts/docker-entrypoint.sh as its CMD. Cloned with core.autocrlf=true (the Git for Windows default), the script was checked out with CRLF, the kernel read the shebang as /bin/bash\r, and the container died immediately with exec: /app/scripts/docker-entrypoint.sh: not found (exit 127). A .gitattributes now pins *.sh to LF at checkout on every platform.

[3.0.1] - 2026-08-06​

Fixed — interactive Google login on global installs / stdio clients (#27)​

The interactive login could not be completed on a global (npm install -g) install driven by a stdio MCP client (e.g. Claude Desktop): the visible-browser login only shipped as an internal script with no exposed command, so users were forced to trigger it through the in-client MCP tool — where the client's tool-call timeout kills the up-to-10-minute Google login before it can finish. get_health reporting headless: true (the runtime query default, unrelated to the setup browser) and a bare headless: false param being silently dropped compounded the confusion.

  • New notebooklm-mcp-setup-auth bin and a notebooklm-mcp setup-auth subcommand (also de-auth, accounts, help) so the interactive login runs as a first-class terminal command — no clone, no in-client tool, no timeout.
  • setup_auth / re_auth now accept a top-level headless boolean (inverse alias of show_browser) instead of silently ignoring it.
  • Docs updated to tell stdio-client users to run the login in a terminal rather than asking the assistant to "log me in".

[3.0.0] - 2026-07-30​

Changed — major refactor: dual transport (internal RPC API + DOM fallback)​

The data plane no longer scrapes the browser DOM for every operation. It now drives NotebookLM's internal batchexecute RPC API — the same JSON-RPC endpoint the web app itself calls — with the Playwright browser retained as an automatic fallback and for login / auto-reauth. Both paths ship permanently: the client tries RPC first and falls back to DOM automation if an internal endpoint shifts, so a future UI rebrand degrades gracefully instead of breaking.

  • Immune to UI rebrands — the operations that broke in the "Gemini Notebook" rebrand (list / create / delete / sources / ask / Studio) run over stable internal ids, independent of the DOM.
  • 10-100× faster — list notebooks ~1 s (was ~30 s), generate a report ~13 s and a data table ~18 s (was minutes). Adding a URL source returns the source id directly instead of polling up to 60 s.
  • More correct — list_notebooks returns every notebook (the DOM path missed notebooks hidden by homepage filter/view state); ask returns structured citations (source id + citation number + cited passage) parsed from the streaming query response rather than reconstructed from DOM highlights.
  • Force the transport with NOTEBOOKLM_TRANSPORT=dom (global DOM) or hot-patch a rotated RPC id via NOTEBOOKLM_RPC_OVERRIDES (JSON map) without a release.
  • Download tries the RPC binary fetch, detects the HTML viewer page Google serves for Studio media, and falls back to the DOM downloader automatically.

Added — 5 new tools (RPC only)​

  • share_notebook — read/set a notebook's public-link status and list collaborators.
  • generate_study_aid — generate flashcards or a quiz from the sources.
  • generate_mind_map — generate and save an interactive mind map.
  • manage_labels — list / create / rename / delete source labels.
  • research_sources — run NotebookLM's web research (fast or deep) to discover and optionally import new sources.

All five are exposed over MCP, the HTTP REST API, and src/index.ts dispatch.

Notes​

  • No breaking changes to existing tool names or the REST surface; the RPC migration is transparent. The major bump reflects the architectural change and the new permanent dual-transport contract.

[2.3.0] - 2026-07-30​

Fixed — the "Gemini Notebook" rebrand rebuilt the whole UI; create / list / rename / delete / sources / Studio all repaired (#23, #21)​

Google's rebrand of NotebookLM to "Gemini Notebook" rebuilt the homepage and Studio DOM, breaking most browser operations. All of the following were diagnosed and re-verified end-to-end against a live authenticated account:

  • create_notebook — the SPA polls the network forever, so waitUntil: 'networkidle' never fired (30 s hang). Switched to domcontentloaded; also dismiss the first-run onboarding overlay.
  • rename in create_notebook — the title is now input.title-input inside <editable-project-title> (no aria-label / placeholder), and the auto-opened add-source dialog blocked it. Dismiss the overlay, then fill the input. Verified the name persists.
  • list_notebooks — project-{UUID} ids and /notebook/<id> hrefs only render in grid view; a list-view account returned 0 on a non-empty account. Force grid view before scraping (removes the slow list-view click-through and its recency-resort race, #21).
  • delete_notebooks — the tile walk overshot to a shared container and deleted the wrong notebook while mis-reporting it. Target the enclosing <project-button> tile, scope the confirm to the dialog, and verify the id is gone before reporting success (data-loss safety).
  • add_source (URL) — false "upload failed" for URL sources that fetch server-side (20-40 s); the check waited ~5 s. Now polls up to 60 s.
  • generate_content (Studio) — Studio became a grid of generation cards. Trigger by the card's language-independent mat-icon (forced past a hover tooltip), handle the per-type preset / customization dialog, and detect completion via a new <artifact-library-item>. All six types verified: audio, report, data table, presentation, infographic, video.
  • list_content / download_content — rewritten on the artifact-library model; download opens the artifact's ⋮ menu → "Télécharger" (NotebookLM now serves audio as .m4a).

Also extends NOTEBOOKLM_UI_LOCALE to fr / en / de / ja. Thanks to @Excauboi (#21) and @KolinEchout (#23) for the reports.

[2.2.1] - 2026-07-25​

Fixed — Workspace accounts whose app lives on notebook.google.com were not recognized (#19)​

NotebookLM is reachable at two hosts: notebooklm.google.com (Google's documented canonical host) and notebook.google.com (the "Gemini Notebook" alias). Which one an authenticated session resolves to is account-dependent — some Workspace tenants land on notebook.google.com (their service-session cookie is bound there). The tool hard-coded notebooklm.google.com in every detection/validation check, so for those accounts it rejected pasted notebook URLs and never detected login success ("session expired server-side" with valid cookies).

All host checks now accept both hosts via a single isNotebookHost / isNotebookUrl module; navigation still targets the documented canonical host and follows Google's redirect. URL parsing also replaces a substring includes() that could false-match an accounts.google.com sign-in URL carrying ?continue=…notebooklm.google.com. Validated end-to-end on a real personal account (interactive login + notebook_ask on live notebooks). Thanks to @kpietkaa, who diagnosed the Workspace host-scoping and drove the revised patch.

Fixed — notebook listing was ~30s slower after the "Gemini Notebook" rebrand​

The rebrand changed the homepage notebook-card title wrapper from a <button> to an <a>. The scrape's initial waitForSelector still targeted the <button>, so on any account that owns notebooks it timed out a full 30s before the (already tag-agnostic) id-based scan found them. Listing was correct but took ~34s; it now resolves in ~5s. The wait now matches the stable project-{uuid}-title id instead of a specific tag.

Fixed — HTTP startup banner showed a hard-coded version​

http-wrapper.ts printed v1.5.3 regardless of the real package version. It now reads package.json at runtime (same as the MCP entrypoint), so the banner can never drift again.

[2.2.0] - 2026-07-11​

Fixed — new-answer detection could time out on a repeated answer text​

waitForLatestAnswer (used by notebook_ask and every Studio content generation flow) identified "the new answer" by hashing its text and comparing against the texts captured before the question was submitted. If NotebookLM answered the new question with the exact same text as an earlier answer in the same notebook (e.g. two questions both answered "Yes", or repeated short/deterministic test prompts), the new answer's hash collided with an already-known one and was skipped as "seen" — the wait then timed out after the full 5-minute budget despite the answer being stable and visible in the page.

Fixed by identifying the new answer via DOM position instead: NotebookLM adds exactly one new .to-user-container per submitted question, so "any container added after a pre-submission baseline count" is immune to text collisions. New countAnswerContainers() helper in page-utils.ts; the position baseline is threaded through notebook_ask, content_generate's chat-fallback path, and the audio overview chat-fallback path. Falls back to the previous hash-based comparison when no baseline is supplied (unchanged behavior for that case). 4 new regression tests reproduce the exact hash-collision scenario and assert the position-based path resolves it.

Root cause identified while reviewing a similar diagnosis in PleasePrompto/notebooklm-mcp#61 — thanks to @macho715 for the writeup; this fix is our own implementation, adapted to this project's architecture.

Fixed — orphan process on stdio disconnect​

The server only handled SIGINT/SIGTERM for graceful shutdown. MCP clients (Claude Desktop, Claude Code, Codex, …) signal disconnect by closing the stdio pipe, not by sending a signal — without a handler for that, the process could stay alive as an orphan with its browser session still open. Now also shuts down on stdin end/close. Inspired by PleasePrompto/notebooklm-mcp#47 (credit: @altristech2025).

Added — Japanese UI locale​

New src/i18n/ja.json, registered alongside fr/en/de. Best-effort translations, not runtime-validated against a live Japanese NotebookLM account — open an issue if any string is off. Inspired by PleasePrompto/notebooklm-mcp#51 (credit: @toshieji), reimplemented against this project's i18n system rather than ported directly (different codebase architecture).

[2.1.1] - 2026-07-11​

Added — Thai locale selectors for notebook_create (#18, partial)​

Added Thai text/aria-label selectors (สร้าง, สร้างใหม่, สร้างโน้ตบุ๊ก) to the "Create notebook" button detection, contributed by @nontakiat. Without this, notebook creation failed on Thai-language NotebookLM accounts, the same class of issue German hit in [2.0.4].

Only the selector addition from #18 is included in this release. The PR's second change — treating a NotebookLM notebook-redirect during text-source upload as success — was not merged: it reports success: true even though the source lands in the wrong (auto-created) notebook, not the one the caller asked for. See the PR discussion for details.

[2.1.0] - 2026-06-26​

Added — note_list and note_get MCP tools (#17)​

Two new Studio-panel tools, contributed by @he0xwhale:

  • note_list (legacy alias list_notes) — scrolls the Studio panel to trigger lazy-load of older notes, then scrapes the note cards. Returns { id?, title, details? } per note (timestamp lives in details).
  • note_get (legacy alias get_note) — opens a note by title (or ID) and serializes its DOM to markdown. Includes a structural walker over element-list-renderer that preserves tables (with header row), ordered/unordered lists, headings, code blocks, and inline bold/italic/code/link spans. Falls back to the ProseMirror editor and then to the legacy text-only extractor if the structural walker comes back empty.

Both tools are exposed over HTTP at POST /content/notes/list and POST /content/notes/get. Selectors use the durable artifact-library-note + artifact-title pattern, consistent with the id-pattern convention used elsewhere in the scraper.

Minor release: additive only, no behavioural change to any existing tool.

[2.0.4] - 2026-06-20​

Added — German UI selectors (closes #14 deferred item)​

The selector pipeline used to iterate only fr + en, so NotebookLM accounts running in German (any selector-based action — add_source, generate_content, the Studio tab, the source-type chips) would silently fall through and report "Could not find …" errors.

  • New src/i18n/de.json with the full string set (tabs, buttons, source types, content types, status, errors, placeholders, content options).
  • 'de' registered as a SupportedLocale so SelectorBuilder, tAll() and every dialog/chip lookup now emit German variants alongside FR/EN.
  • New src/__tests__/i18n.test.ts asserts every locale ships the same leaf-key shape as the EN reference — catches the "new locale missing a translation key" regression class.

The German strings are best-effort and not runtime-validated against a live German NotebookLM account — open an issue / PR if any string is off and we'll fix it in a patch.

[2.0.3] - 2026-06-20​

Security — vault.batch path-traversal hardening (issue #15)​

The vault_batch tool (and the equivalent POST /batch-to-vault HTTP endpoint) previously passed the caller-supplied vault_dir straight to path.resolve() + fs.mkdir(), with no containment check. A caller (or a prompt-injected LLM driving the MCP) could therefore create directories and write .md / .json files anywhere the server process can write.

The fix is opt-in to stay backward-compatible:

  • NOTEBOOKLM_VAULT_ROOT env var — when set, every vault_dir is resolved relative to this root and the resolved path must stay inside realpath(NOTEBOOKLM_VAULT_ROOT). Any attempt to escape via an absolute path or a .. segment is rejected with a clear error.
  • slug_prefix is now sanitized — path separators (/, \), .. sequences and NUL bytes are stripped, and the prefix is capped at 64 chars, so a caller can no longer smuggle path traversal through the filename component.

Existing users who do not set NOTEBOOKLM_VAULT_ROOT see no behaviour change; the slug_prefix sanitization is always on but only strips bytes that were never valid filename characters anyway.

Fixed — CI security audit no longer fails on dev-only advisories​

The Security job now runs npm audit --omit=dev --audit-level=moderate. Vulnerabilities reported in the jest / babel toolchain (which never ship in the published package) no longer turn the main branch red while letting production-dep regressions through. Also bumps hono override to ^4.12.25 to clear GHSA-88fw-hqm2-52qc (CORS wildcard with credentials) and the related cluster of hono advisories.

[2.0.2] - 2026-06-20​

Fixed — canonical tool names use _ instead of .​

The v2 canonical tool names were a dot-separated tree (notebook.ask, source.add, …), but the MCP / Anthropic tool-name pattern is ^[a-zA-Z0-9_-]{1,64}$ — dots are not allowed. Claude Code silently rewrote the dots, masking the problem, but Claude Desktop validates remote tool definitions strictly and rejected the whole server with:

tools.<n>.FrontendRemoteMcpToolDefinition.name: String should match pattern '^[a-zA-Z0-9_-]{1,64}$'

Every canonical name now uses _ as the namespace separator: notebook.ask → notebook_ask, source.add → source_add, vault.batch → vault_batch, and so on across all 9 namespaces. This is what Claude Code already displayed, so its users see no change. The legacy flat names remain accepted as aliases, and a regression test now asserts every advertised name matches the pattern.

[2.0.1] - 2026-05-27​

Fixed​

  • add_source for URL sources — the 2026 NotebookLM add-source dialog redesign broke URL uploads on three fronts: the "Websites" tile is now an Angular Material chip (mat-chip / mat-mdc-chip), the URL textarea ships aria-label="Enter URLs" instead of a placeholder containing "URL", and the verification path's Escape keypress was landing mid-ingest and canceling the add. The dialog also closes asynchronously now, so the baseline source count was being captured after the new row already appeared and the count-based detection never tripped. Threading the pre-baseline count from addSource() and waiting for the dialog to settle before verification fixes the false-negative. (Reported by @assistantthatskywalker in #14.)
  • Answer wait timeout 2 min → 5 min in waitForLatestAnswer — long NotebookLM answers with many citations regularly exceed 2 minutes and were timing out as false negatives. (Cherry-picked from @LScelza-synapse's fork.)
  • Spanish profile-lock locale detection — added "Abriendo en una sesión existente" to the regex that triggers the singleton-profile fallback, so Spanish-locale Chrome installs no longer hard-fail on a locked profile.

Security​

  • Override brace-expansion@^5.0.6 (scoped to the ^5.0.0 range so the older 1.x line in eslint deps is untouched) — fixes GHSA-jxxr-4gwj-5jf2.
  • Override qs@^6.15.2 — fixes GHSA-q8mj-m7cp-5q26 in the express / body-parser transitive chain.

[2.0.0] - 2026-05-14​

Changed — tool names are now a namespaced tree​

Every tool was renamed from a flat snake_case list to a navigable namespace_action tree (9 namespaces): notebook_*, library_*, session_*, source_*, content_*, note_*, auth_*, server_*, vault_*. For example ask_question → notebook_ask, add_source → source_add, list_sessions → session_list, get_health → server_health, batch_to_vault → vault_batch. tools/list now advertises only the canonical names. This aligns with MCP naming best practice and is the reason for the major version bump.

Note: 2.0.0–2.0.1 shipped these names with a . separator (notebook.ask), which the MCP tool-name pattern forbids; 2.0.2 corrected the separator to _. See the [2.0.2] entry.

Backward compatible — nothing breaks. Both the stdio server and the HTTP proxy still accept the legacy flat names as aliases: the dispatch layer normalises any accepted name (canonical or legacy) before routing. Existing scripts, Claude Code / Cursor / n8n configs, and batch jobs that hard-code the old names keep working unchanged. New integrations should use the canonical names — the legacy aliases are supported but no longer advertised.

The legacy ↔ canonical mapping lives in one place: src/tools/tool-names.ts.

Added​

  • MCP annotations on every tool — accurate readOnlyHint / destructiveHint / idempotentHint / openWorldHint / title per tool, from a central table in src/tools/index.ts.
  • MCP outputSchema on every tool — the shared { success, data?, error? } result envelope every handler returns. The stdio dispatch now also returns structuredContent matching that schema, not just text.
  • smithery.yaml + mcpb/ bundle source — the server is now published on the Smithery registry.

Security​

  • Resolve two new transitive advisories pulled in via @modelcontextprotocol/sdk@1.26.0: fast-uri ≤3.1.1 (high — path traversal / host confusion, GHSA-q3j6-qgpj-74h6, GHSA-v39h-62p7-jpjc) and hono ≤4.12.17 (moderate — several, incl. cache cross-user leakage GHSA-p77w-8qqv-26rm). Added overrides for fast-uri ^3.1.2 and hono ^4.12.18; npm audit reports 0 vulnerabilities.

[1.7.9] - 2026-05-06​

Security​

  • Resolve transitive moderate XSS advisory GHSA-v2v4-37r5-5v8g in ip-address ≤10.1.0 (reachable via @modelcontextprotocol/sdk@1.26.0 → express-rate-limit@8.x → ip-address). Added overrides: { "ip-address": "^10.2.0" } to package.json to force the patched version regardless of upstream re-resolution. npm audit now reports 0 vulnerabilities. CI security job (audit gate) was failing on 1.7.8 because of this advisory; this release unblocks it.

[1.7.8] - 2026-05-06​

Fixed (this time properly tested at runtime, MCP attached)​

add_source returned Timeout waiting for source processing for uploads that actually succeeded — false negative.

Verified in a live MCP session against my own NotebookLM account. After the upload, list_content immediately reported the source as status: "ready" while add_source reported success: false. Root cause traced in waitForSourceProcessing: NotebookLM 2026 keeps the upload dialog open after a successful upload (it lets the user chain more uploads), and the count-based success detection in the polling loop was gated behind if (!dialogVisible) — so the loop ran for the full 90s timeout while the dialog stayed open, and the early success signal (the source-list count growing by 1) was never tripped. Fix: lift the count-based detection out of the dialog-state branch and run it on every poll cycle. As soon as the source-list count grows, return success regardless of whether NotebookLM has dismissed the dialog or not.

Packaging: dist/index.js and dist/stdio-http-proxy.js were published in mode 644 (no executable bit) since at least 1.7.3.

The shebang #!/usr/bin/env node was present, but without the executable bit npm-published archives only worked on systems where the npm install step happened to re-chmod (e.g. the bin symlink machinery). Sandbox bash environments that don't re-chmod silently failed with Permission denied. Fix: new build:chmod-bin step in the build script that explicitly chmod-755s both binaries before they get tarballed by npm publish. Verified locally — ls -la dist/index.js now shows -rwxr-xr-x after npm run build.

[1.7.7] - 2026-05-06​

Defensive (not yet validated runtime)​

add_source Could not find "Add source" button — possibly fixed, definitely better diagnosed.

Reported by an end-user agent against a freshly-created notebook. Root cause is uncertain without runtime visibility from a Claude Code session — but two things are now in place:

  • Selectors broadened to cover the empty-notebook layout NotebookLM shows for fresh notebooks (centered "Upload sources" / "Choose files" CTA in EN, "Téléverser" / "Importer" in FR), in addition to the per-panel "+" button. Both aria-label and visible-text variants are tried.
  • Error is now actionable. Previously the failure surfaced as just Could not find "Add source" button. The error now embeds a JSON dump of what the page contains: URL, document title, total button count, and up to 25 buttons with their aria-label, text content, mat-icon presence, class list, and visibility flag. Plus the path of the debug screenshot. So when this error reaches the calling agent, it can either pick the right selector itself or paste the dump back to a maintainer for a precise fix.

I do not have runtime access to a NotebookLM session in this environment, so I cannot confirm the broadened selectors actually resolve the error. If the issue persists, the enriched payload is designed to make the next iteration definitive instead of speculative.

[1.7.6] - 2026-05-06​

Fixed​

Two bugs in tools that 1.7.4 exposed without runtime validation. Honest disclosure: I shipped both create_notebook and delete_notebooks_from_nblm in 1.7.4 without exercising them against a live NotebookLM session — only checked that the handler methods existed and the dispatch was wired. Both had latent issues that surfaced as soon as a real user ran them.

create_notebook​

Two regressions, both reported by an end-user agent that ran the tool against the live UI:

  • The tool returned a transitional URL like https://notebooklm.google.com/notebook/creating/c instead of the final notebook URL. Root cause: the waitForURL regex (/notebook\//) matched the transitional URL immediately rather than waiting for the redirect to the UUID-based URL. Fix: require a full v4 UUID in the URL pattern, then wait for networkidle before reading the URL.
  • The name parameter was silently ignored — every notebook came back as "Untitled notebook". Root cause: the rename used a generic [contenteditable="true"], .notebook-title, h1 selector, took the first match (often the wrong element), typed via Control+a+type, and never verified the result. Fix: target ordered, specific selectors (aria-label / placeholder containing "title"/"Untitled"), use fill for <input> and evaluate+keystroke for contenteditable, then read back the title to verify the rename took. The tool now returns name_applied: boolean and actual_name: string so the caller knows whether to retry via the UI.

delete_notebooks_from_nblm​

The handler used button[aria-labelledby*="project-"] as its primary selector — the same pattern that broke list_notebooks_from_nblm and was rewritten in 1.7.5. So the delete tool was very likely non-functional on the current NotebookLM DOM. Fix: locate each notebook tile through its stable id="project-{UUID}-title" element (a page.evaluate() walks up to the nearest card-like ancestor and stamps it with data-nblm-target-card), then resolve the menu button inside that marked card. Same id-pattern strategy that already powers list_notebooks_from_nblm.

Process note​

I am updating the project memory to require either a runtime smoke test or a full read of the handler implementation (selectors, waitForURL, post-action verification) before exposing any handler as a new MCP tool. Skipping that audit in 1.7.4 cost two extra releases.

[1.7.5] - 2026-05-06​

Three bugs reported by an end-user agent against 1.7.2–1.7.4. All three are fixed in this release.

Fixed​

#1 — /plugin marketplace update + /reload-plugins did not actually upgrade the running MCP server (cache-poisoning by npx -y without a version pin).

mcpServers.notebooklm.args was ["-y", "@roomi-fields/notebooklm-mcp"]. Without a version specifier, npx reused the existing _npx/<hash>/ cache regardless of which version the marketplace just refreshed to. Symptom: plugin manifest showed 1.7.4 but the live MCP server was still 1.7.2 and rejected newly-added tools (create_notebook etc.) with Unknown tool.

Fix: pin the version inside the args ("@roomi-fields/notebooklm-mcp@1.7.5"). scripts/sync-version.mjs now propagates the package version to both the manifest's version field AND the npx pin in mcpServers.notebooklm.args. The CI version:check validates both. From now on every release ships with a manifest that points npx at exactly the version the marketplace just resolved to.

#2 — list_notebooks_from_nblm returned every notebook with name: "Notebook".

The Strategy 1 selector (button[aria-labelledby*="project-"]) no longer matches the current NotebookLM DOM, so it returned 0. The Strategy 2 fallback then scraped UUIDs out of the raw HTML and hardcoded name: 'Notebook' for each — making the API report N notebooks all named "Notebook" instead of an empty result or real titles.

Fix: replace the brittle button selector with a DOM-agnostic page.evaluate() that walks every element whose id matches project-{UUID}-title and reads its textContent. The id-based naming is intrinsic to NotebookLM's data model and stable across UI rewrites. Strategy 2 (raw HTML fallback) now returns name: '' for ids it can only discover from HTML — never the misleading hardcoded placeholder.

#3 — get_health.current_account reported the previous account after re_auth succeeded with a different Google account.

handleReAuth ran de-auth + performSetup() (manual browser login) but never touched the current-account.txt marker. Since performSetup() does not know which Google account the user picks during browser login, the marker stayed stale and get_health kept returning the old email.

Fix: new AccountManager.clearCurrentAccountId() method, called after a successful re_auth. get_health now truthfully omits current_account rather than returning a wrong value. Detecting the new account from the post-login browser state is a richer follow-up — for now we choose honest absence over misleading presence.

[1.7.4] - 2026-05-06​

Added​

Two previously orphaned NotebookLM management tools are now exposed via MCP:

  • create_notebook({ name?, show_browser? }) — creates a brand-new empty notebook directly in NotebookLM (no pre-existing share URL required, unlike add_notebook which only registers an already-created notebook into the local library). Returns the freshly minted notebook_url and notebook_id so the caller can chain add_source, ask_question, etc.
  • delete_notebooks_from_nblm({ notebook_ids[], show_browser? }) — bulk-delete notebooks at the NotebookLM UI level (not just from the local library). Returns { deleted, failed, message }. Pair it with list_notebooks_from_nblm to discover IDs.

Internal​

Audited the full tool surface: 30 ToolHandlers methods are now aligned 1:1 with 30 tool definitions and 30 dispatch cases. No more declared-but-unwired tools, no more handlers without a tool definition.

[1.7.3] - 2026-05-06​

Fixed​

  • list_notebooks_from_nblm MCP tool was declared in the tool list but missing from the dispatch switch in src/index.ts, so calling it returned Unknown tool. Added the missing case → the tool now resolves and runs handleListNotebooksFromNblm.

Known orphans (not fixed in this patch)​

  • handleCreateNotebook and handleDeleteNotebooksFromNblm exist as ToolHandlers methods but are not exposed via either a tool definition or a switch case. They are unreachable today and have no effect on user-visible behaviour. A follow-up will decide whether to expose them or remove them.

[1.7.2] - 2026-05-06​

Added​

Claude Code plugin manifest + cross-file version sync:

  • New .claude-plugin/plugin.json at the repo root — Claude Code plugin manifest (single source of truth for plugin metadata, including mcpServers config). Lets the roomi-fields/claude-plugins aggregated marketplace source this project directly via source: github instead of maintaining a duplicated wrapper.
  • New scripts/sync-version.mjs + npm scripts version:sync / version:check — propagates package.json.version to the plugin manifest, website/docusaurus.config.ts (softwareVersion), and the README hero line. Uses an in-place regex replace for the plugin manifest so prettier-controlled formatting is preserved across releases (no phantom drift in CI).
  • README: new "Option 0 — Claude Code marketplace" Quick Start section pointing at the aggregated marketplace one-liner.

Changed​

  • .github/workflows/release.yml: added a "Verify version sync" step before build/publish — so a release can never ship with a stale plugin manifest or website badge.

Fixed​

  • 1.7.1 was tagged but the publish failed when the version-sync check rejected a prettier-induced JSON-formatting drift. This release supersedes that aborted tag.

[1.7.0] - 2026-05-06​

Added​

batch_to_vault is now a first-class MCP tool — no HTTP server required:

  • New MCP tool batch_to_vault exposed alongside ask_question, add_source, etc. Same parameters and output as the HTTP endpoint (questions[], vault_dir, notebook_id/notebook_url, slug_prefix, source_format, sleep_between_ms, session_id). MCP clients (Claude Code, Cursor, Codex…) can now run a batch directly without starting npm run start:http or hitting localhost:3000.
  • Shared helper runBatchToVault in src/utils/vault-writer.ts — single source of truth for the batch loop. Both the MCP tool (ToolHandlers.handleBatchToVault) and the HTTP /batch-to-vault endpoint call it with the same askQuestion injection. Eliminates the previous duplicated loop in http-wrapper.ts.

Changed​

  • POST /batch-to-vault is now a thin wrapper over ToolHandlers.handleBatchToVault. Behaviour and response shape are unchanged; the implementation is just no longer duplicated.

[1.6.0] - 2026-05-04​

Added​

RTFM integration — NotebookLM as one-shot ingestion, RTFM as retrieval layer:

  • POST /batch-to-vault endpoint: runs a list of questions and writes each answer as a markdown file (YAML frontmatter + answer + sources block) and a JSON sidecar in a vault directory, ready for ingestion by RTFM or any markdown vault tool (Obsidian, Foam, Dendron). Removes the 50-queries/day quota and 10–30s latency for repeat queries on academic / SOTA workflows.
  • src/utils/vault-writer.ts: formatAnswerMarkdown, formatAnswerJson, makeSlug helpers with NFD-normalised slugs and YAML escape.
  • nblm-answer-v1 JSON Schema published at the canonical, immutable URL https://schemas.roomi-fields.com/nblm-answer-v1.json (served as application/schema+json with open CORS, JSON-LD Dataset entry for AI-crawler discovery, 24h immutable cache). Source-of-truth copy in schemas/nblm-answer-v1.json.
  • New documentation page 14-RTFM-INTEGRATION.md — full pattern, schema reference, RTFM wiring, recommended layout for academic batches, question-generation prompt.
  • README: RTFM listed as the 4th Integration Option alongside MCP, HTTP REST, and Docker.

Changed​

  • Root / endpoint now lists batch_to_vault in its endpoint registry.

[1.5.9] - 2026-04-22​

Fixed​

  • Restore mcpName field in package.json (accidentally dropped during the 1.5.8 release). This field is required by the official MCP Registry to verify npm-package ownership; without it the registry refuses to update server metadata. v1.5.9 re-enables registry updates for io.github.roomi-fields/notebooklm-mcp.

[1.5.8] - 2026-04-19​

Fixed​

NotebookLM UI adaptations for 2026 (PR #5 by @KhizarJamshaidIqbal):

  • sanitizeResponseText(): strip leaked Material-icon labels (more_vert, more_horiz) and adjacent citation artifacts from answer text
  • ensureDiscussionReady() / navigateToDiscussion(): recover the Discussion panel before asking when the chat composer has been unmounted or disabled
  • Count-based source detection: use .single-source-container row counts instead of relying on source names, so URL sources (which surface as page titles, not URLs) are still detected correctly
  • New UI selectors for the "Add source" flow (.cdk-overlay-pane, role=menu, role=menuitem)
  • Chat-input detection now skips visible-but-disabled composers
  • Regression tests added in src/__tests__/{page-utils,browser-session,content-manager}.test.ts

Added​

Doctor scripts (PR #6 by @KhizarJamshaidIqbal):

  • npm run doctor:basic — repo + build artifacts + Node engine check
  • npm run doctor:http — /health, optional /content and /ask verification (pass --notebook-url)
  • .env.example documenting the project's config knobs

Security​

  • Credentials, personal paths, test-account emails, and real notebook UUIDs scrubbed from archived scripts and docs (PR #4 by @KhizarJamshaidIqbal and follow-up commits)
  • Two tracked debug PNGs with literal Windows paths in their filenames removed from the index
  • npm audit fix applied: hono 4.12.14, @hono/node-server 1.19.14, express-rate-limit 8.3.2, handlebars 4.7.9, brace-expansion, flatted, ip-address bumped to patched versions (no direct-dependency changes)

[1.5.7] - 2026-02-26​

Fixed​

Citation Extraction:

  • Fixed .highlighted selector: use .highlighted class (not i.highlighted) for citation excerpt extraction
  • Scoped citation button detection to last .to-user-container .message-text-content container — fixes multi-question sessions picking up buttons from all previous answers

Changed​

Docker (PR #1 by @JulienCANTONI):

  • Dockerfile now builds TypeScript from source inside the container instead of copying pre-built dist/
  • Multi-stage: install all deps → build → prune devDeps → install browser (cleaner CI/CD)
  • .dockerignore updated to allow src/ and tsconfig*.json for in-container build
  • docker-entrypoint.sh now supports PORT env var (Render PaaS compatibility)

Security​

  • Updated @modelcontextprotocol/sdk to fix vulnerabilities
  • Updated qs dependency to fix prototype pollution vulnerability

[1.5.6] - 2026-02-15​

Fixed​

Citation Extraction (Major Rewrite):

  • Rewrote citation excerpt extraction using page.evaluate() string expressions — eliminates all ElementHandle stale reference errors
  • Source names extracted from span[aria-label] in a single DOM scan (100% reliable)
  • Source excerpts extracted by clicking each citation → reading i.highlighted + parent .paragraph for full passage context
  • Added Escape dismiss between each citation click to prevent stale highlight contamination
  • Improved from ~50% to 97% success rate (30/31 citations in real-world test)
  • Removed all debug screenshots and verbose logging from production code

Authentication & Session Recovery:

  • Fixed false "Authenticated" at startup: verifyWithBrowser() now does real browser navigation to NotebookLM to check if Google session is truly valid (not just local cookie expiry dates)
  • Fixed mid-session SESSION_EXPIRED: auto-reauth now uses AutoLoginManager with stored credentials (fills email + password automatically), falls back to manual performSetup() if needed
  • Fixed profile sync direction bug: after performSetup(), uses syncMainToAccount() (not the reverse) to avoid overwriting fresh auth with stale account cookies
  • Fixed setup-auth --force flag to bypass cookie check and always open browser

Port Management:

  • Added EADDRINUSE handling: detects ghost processes on port 3000, attempts auto-kill, provides clear error message if port cannot be freed

Added​

Chrome Profile Auto-Sync:

  • syncProfileToMain(accountId): copies account-specific state/profile → main chrome_profile/
  • syncMainToAccount(accountId): reverse sync after interactive re-auth
  • Both called automatically at the right points in startup and mid-session flows

Startup Browser Verification (Step 4):

  • New verifyWithBrowser() method launches headless browser, navigates to NotebookLM, checks actual URL
  • If redirected to accounts.google.com → triggers auto-reauth with AutoLoginManager (automatic) or performSetup() (manual fallback)

[1.5.5] - 2026-01-31​

Fixed​

Authentication State Path Mismatch (Critical Bug):

  • Fixed /health endpoint and session recovery checking wrong path for authentication cookies
  • Root cause: Multi-account system saves cookies to accounts/{id}/browser_state/state.json but code was checking legacy global path CONFIG.browserStateDir/state.json
  • Fixed in 4 files:
    • src/startup/startup-manager.ts - Now uses account-specific state path at startup
    • src/tools/index.ts - Health endpoint checks account-specific state with fallback
    • src/session/shared-context-manager.ts - Context recreation loads from correct path
    • src/session/browser-session.ts - Session recovery uses account-specific state
  • All files maintain fallback to legacy AuthManager path for backward compatibility

Health Endpoint Improvements:

  • Added current_account field to health response showing active account email
  • Added accountCheckDone flag to ensure proper fallback to AuthManager in tests

Added​

Windows Startup Scripts:

  • scripts/start-server-hidden.vbs - VBScript to launch HTTP server silently at Windows startup
  • scripts/stop-server.bat - Batch script to stop running server processes
  • scripts/mcp-proxy-hidden.ps1 - PowerShell script for hidden window MCP proxy

Claude Code MCP Integration:

  • Hidden window configuration for MCP stdio proxy in Claude Code projects
  • Uses powershell -WindowStyle Hidden to avoid shell window popup on each tool call
  • Example .mcp.json configuration in documentation

[1.5.4] - 2026-01-07​

Fixed​

Citation Source Format (Critical Bug):

  • Fixed source_format parameter not being passed from HTTP /ask endpoint to handler
  • The parameter was defined in the schema but never extracted from req.body in http-wrapper.ts
  • All source formats now work correctly: none, inline, footnotes, json, expanded

Citation Extraction:

  • Updated CITATION_SELECTORS for current NotebookLM DOM (January 2026):
    • Primary: button.citation-marker, button.xap-inline-dialog.citation-marker
    • Fallback selectors for backwards compatibility
  • Updated TOOLTIP_SELECTORS for source text extraction:
    • Primary: i.highlighted, .paragraph i.highlighted
  • New extractSourceFromElement() function extracts both:
    • sourceName: from aria-label attribute (e.g., "17: Filename.pdf")
    • sourceText: from hover tooltip content

Docker/noVNC:

  • Fixed Xvfb startup error: _XSERVTransmkdir: ERROR: euid != 0
  • Added /tmp/.X11-unix directory creation with proper permissions (1777) in Dockerfile
  • Changed exposed port from 5900 (raw VNC) to 6080 (noVNC web interface) in docker-compose.yml

Added​

  • E2E tests for source_format parameter: [T15], [T16], [T17] in tests/e2e/tests/03-ask.test.ts

[1.5.3] - 2026-01-05​

Added​

Docker Deployment with noVNC:

  • noVNC integration for visual browser authentication in Docker
  • Xvfb + x11vnc + websockify for headless display server
  • Port 6080 for web-based VNC access
  • scripts/start-vnc.sh - VNC services startup script
  • scripts/docker-entrypoint.sh - Container entrypoint combining VNC + Node.js
  • NAS deployment support (Synology, QNAP) with export/import workflow

Bulk Import Endpoint:

  • POST /notebooks/import-from-scrape - Bulk import notebooks from NotebookLM scrape
  • Options: notebook_ids (array) to filter, auto_discover (boolean) for AI metadata
  • Scrapes all notebooks from account and adds them to library in one call

New Configuration Options:

  • browserChannel config: chromium (default) or chrome
  • ENABLE_VNC environment variable for Docker
  • NOVNC_PORT environment variable (default: 6080)

Fixed​

Docker Compatibility:

  • Fixed locale configuration to use CONFIG.uiLocale instead of hardcoded en-US
  • Fixed browser channel for Docker (chromium vs chrome)
  • Fixed patchright browser installation in Dockerfile
  • Fixed show_browser parameter passing to performSetup
  • Added Docker-specific Chrome flags (--no-sandbox, --disable-setuid-sandbox, etc.)
  • Fixed page.goto timeouts with waitUntil: 'domcontentloaded'

HTTP Server:

  • Added root endpoint (/) returning API info

[1.5.2] - 2026-01-01​

Added​

Notebook Scraping from NotebookLM:

  • New list_notebooks_from_nblm tool to scrape real notebooks from NotebookLM homepage
  • Uses correct button selectors (button[aria-labelledby*="project-"]) matching NotebookLM's actual HTML structure
  • Returns notebook IDs, names, and URLs for all notebooks on the account

Bulk Notebook Deletion:

  • New delete_notebooks_from_nblm tool for bulk notebook deletion
  • New DELETE /notebooks/bulk-delete HTTP endpoint for batch deletion
  • Support for protected notebook IDs that won't be deleted
  • Progress tracking during deletion operations

Fixed​

Tool Description Builder:

  • Fixed TypeError: Cannot read properties of undefined (reading 'map') in buildAskQuestionDescription
  • Added defensive checks with optional chaining for active.topics and active.use_cases
  • Fallback values when notebook metadata is incomplete

[1.5.1] - 2026-01-01​

Added​

Multilingual UI Support (i18n):

  • New internationalization system for NotebookLM UI selectors
  • Support for French (fr) and English (en) UI languages
  • NOTEBOOKLM_UI_LOCALE environment variable to set UI language
  • Locale files in src/i18n/ with translated selectors for all UI elements
  • SelectorBuilder class for generating bilingual selectors
  • tAll() function to get translations in all supported languages
  • Documentation: docs/ADDING_A_LANGUAGE.md for adding new languages

Language/Account Switching:

  • New scripts/switch-account-language.sh script for automated language switching
  • Automated Chrome profile cache deletion and re-authentication
  • Syncs new profile to main profile after language change
  • Visual verification with --show flag

E2E Test Infrastructure:

  • Complete E2E test suite with 76 tests (75 pass, 1 skip)
  • Test categories: Health, Notebooks, Ask, Sessions, Sources, Content, Notes, CRUD, Errors
  • Support for QUICK mode (55 tests) and FULL mode (76 tests)
  • Multi-account test configuration with per-account notebook URLs
  • Test tracking in tests/e2e/TRACKING.md

Changed​

  • Updated all test files to use currentNotebooks from config for proper account isolation
  • Jest configuration updated for ESM support with --testPathPatterns

[1.4.2] - 2025-12-29​

Removed​

Fake Content Generation Features:

  • Removed generate_content endpoint for FAQ, Study Guide, Briefing Doc, Timeline, and Table of Contents
  • These features were NOT real NotebookLM integrations - they were just sending prompts to the chat
  • The only REAL content generation NotebookLM supports is Audio Overview (podcast)
  • Updated all documentation to honestly reflect actual capabilities

What was fake:

  • briefing_doc - Was just asking NotebookLM chat to generate a summary
  • study_guide - Was just asking NotebookLM chat to create study materials
  • faq - Was just asking NotebookLM chat to generate FAQs
  • timeline - Was just asking NotebookLM chat to create a timeline
  • table_of_contents - Was just asking NotebookLM chat to create a TOC

What is real:

  • Audio Overview generation - Uses NotebookLM's actual podcast feature
  • Audio download - Downloads the real generated audio file
  • Q&A with citations - Uses NotebookLM's actual chat with source citations
  • Source management - Uses NotebookLM's actual source upload features

[1.4.1] - 2025-12-29​

Fixed​

Source Upload UI Compatibility:

  • Fixed URL source upload with textarea support (NotebookLM uses textarea, not input)
  • Fixed YouTube source upload with textarea support
  • Added French UI locale support (placeholders: "Collez des liens", buttons: "Insérer")
  • Improved fallback detection for input/textarea elements in dialogs
  • Added debug logging for unrecognized dialog elements

Audio Download Improvements:

  • Added navigation to Audio Overview panel before attempting download
  • Added Material Design icon selectors (download, file_download, get_app)
  • Added fallback to extract audio source URL directly from audio element
  • Improved debug logging for download button detection

Changed​

Updated UI selectors for NotebookLM:

  • Updated navigateToStudio() with new Material Design tab selectors
  • NotebookLM now uses mdc-tab / mat-mdc-tab Angular Material components
  • Added tab structure: Sources | Discussion | Studio

Added​

Comprehensive E2E Test Suite:

  • Full E2E test coverage: 22/22 endpoints tested and passing
  • PowerShell-based test runner: tests/e2e/run-e2e-tests.ps1
  • Individual test scripts for each endpoint
  • WSL-to-Windows compatible via cmd.exe /c powershell

Fully Functional Features (All Verified E2E):

  • ✅ ask_question - Q&A with source citations
  • ✅ list_notebooks, select_notebook, search_notebooks - Library management
  • ✅ list_content - View sources and existing artifacts
  • ✅ add_source - Add files, URLs, text, YouTube (all types working)
  • generate_content - Removed in v1.4.2 (was fake - just chat prompts)
  • ✅ generate_audio - Audio overview generation
  • ✅ Session management - Create, reset, close sessions
  • ✅ Authentication - setup_auth, re_auth, de_auth

[1.4.0] - 2025-12-24​

Added​

Content Management Module:

  • New content management system for NotebookLM notebooks
  • Source management:
    • add_source - Add files, URLs, text, YouTube videos, Google Drive documents
    • list_content - View all sources and generated content
  • Content generation:
    • generate_audio - Create podcast-style audio overviews
    • generate_content - Removed in v1.4.2 (was fake - just chat prompts, not real NotebookLM features)
    • download_audio - Download generated audio files
  • Full HTTP REST API support with 6 new endpoints
  • Complete documentation in 10-CONTENT-MANAGEMENT.md

Multi-Account Management:

  • New account management system for multiple Google accounts
  • Account switching: switch_account, list_accounts, get_current_account
  • Dedicated session per account for parallel operations
  • HTTP endpoints for account management
  • Documentation in MULTI_ACCOUNT_SYSTEM.md

[1.3.7] - 2025-12-16​

Added​

Source Citation Extraction:

  • New source_format parameter for ask_question tool with 5 options:
    • none: No extraction (default, fastest)
    • inline: Insert source text inline: [1: "source excerpt..."]
    • footnotes: Append sources at the end as footnotes
    • json: Return sources as separate object in response
    • expanded: Replace markers with full quoted source text
  • New src/utils/citation-extractor.ts module for hover-based citation extraction
  • Extracts source citations by hovering over citation markers in NotebookLM responses
  • No additional API calls required - pure DOM interaction

Fixed​

Citation Detection:

  • Fixed findCitationsByRegex returning empty results - now properly finds DOM elements
  • Added multiple detection strategies: CSS selectors, data attributes, XPath fallback
  • Fixed citation replacement for all occurrence (was only replacing first match)
  • Added descending sort to avoid replacing [1] before [10]
  • Handle NotebookLM's various citation formats:
    • Bracketed: [1], [2]
    • Superscript: text1,2 or text3
    • Stuck together: text123 (citations 1, 2, 3)

[1.3.6] - 2025-11-29​

Changed​

Documentation Restructure:

  • Simplified README.md from 765 to 165 lines (-78%)
  • Extracted roadmap section into dedicated ROADMAP.md file
  • README now focuses on Quick Start with links to detailed docs
  • Better separation of concerns: README for overview, docs/ for details

Added​

  • New ROADMAP.md file with planned features and version history

[1.3.5] - 2025-11-27​

Added​

Quality Tooling & CI/CD:

  • Added Jest testing framework with comprehensive test suite (327 tests)
  • Added ESLint configuration for code quality enforcement
  • Added Prettier for consistent code formatting
  • Added Husky + lint-staged for pre-commit hooks
  • Added GitHub Actions CI workflow with multi-Node version testing (18.x, 20.x, 22.x)
  • Added Codecov integration for coverage tracking
  • Added type-coverage tool (99.01% coverage)

Test Coverage:

  • Unit tests for logger.ts (100% coverage)
  • Unit tests for errors.ts and error types
  • Unit tests for config.ts parsing and validation
  • Unit tests for stealth-utils.ts timing functions
  • Unit tests for cleanup-manager.ts core functionality
  • Unit tests for page-utils.ts selectors and utilities
  • Type system tests for ToolResult<T> discriminated union

Fixed​

  • Fixed Prettier formatting issues in README.md and test files
  • Fixed code style consistency across all source files

[1.3.4] - 2025-11-26​

Fixed​

CLI Scripts:

  • Fixed de-auth.ts CLI script: added missing implementation
  • Improved page load wait logic in authentication flow

Test Reliability:

  • Improved test-auth.ps1 reliability: reduced from 9 tests to 7 focused tests
  • Smart cleanup test that checks auth status before attempting restore
  • Cleanup test now passes regardless of whether manual re-auth is needed
  • Removed strict type validation tests that were testing unimplemented server-side validation

[1.3.3] - 2025-01-26​

Security​

CORS Hardening:

  • Added CORS whitelist configuration via CORS_ORIGINS environment variable
  • Default whitelist allows only localhost origins (ports 3000, 5678, 8080)
  • Blocked external origins no longer receive CORS headers
  • Supports wildcard * for development when explicitly configured

Input Validation:

  • Added Zod schema validation for all HTTP endpoints
  • Validates request bodies with detailed error messages
  • Schemas: AskQuestionSchema, AddNotebookSchema, UpdateNotebookSchema, AutoDiscoverSchema, CleanupDataSchema, ShowBrowserSchema

Express Route Security:

  • Fixed route ordering: static routes (/notebooks/search, /notebooks/stats) now correctly matched before parameterized routes (/notebooks/:id)
  • Prevents route hijacking vulnerabilities

Fixed​

Error Handling:

  • Replaced 30+ empty catch blocks with proper log.debug() logging
  • Improved error visibility for debugging without breaking functionality

Type Safety:

  • Refactored ToolResult<T> to discriminated union type for compile-time safety
  • Fixed ServerState types with proper Browser, SessionManager, AuthManager types
  • Added JSONSchemaProperty type for MCP tool input schemas
  • Added config validation with constraint checking (min/max ranges, positive values)
  • Fixed parseProfileStrategy to avoid unsafe as any type assertions

Added​

Test Coverage (25 → 72 tests, +188%):

  • test-validation.ps1 (18 tests) - Zod schema validation testing
  • test-auth.ps1 (8 tests) - Authentication endpoint testing
  • test-cors.ps1 (10 tests) - CORS configuration testing
  • test-sessions.ps1 (10 tests) - Session management testing
  • Fixed test-errors.ps1 pattern matching for Zod validation messages

[1.3.2] - 2025-01-24​

Added​

Authentication Management:

  • New MCP tool de_auth for secure logout (clears all credentials without re-authenticating)
  • Separation of concerns: de_auth (logout only), re_auth (logout + re-authenticate), setup_auth (first-time)
  • HTTP API endpoints for complete authentication lifecycle:
    • POST /de-auth - Logout and clear credentials
    • POST /re-auth - Re-authenticate with different account
    • POST /cleanup-data - Clean all data (requires confirmation)

HTTP API Feature Parity:

  • Added 7 missing endpoints to achieve 100% parity with MCP stdio tools
  • Authentication: /de-auth, /re-auth, /cleanup-data
  • Notebooks: PUT /notebooks/:id, /notebooks/search, /notebooks/stats
  • Sessions: POST /sessions/:id/reset
  • All 22 endpoints now available via both HTTP REST API and MCP stdio

Documentation:

  • Complete API reference updated with all 22 endpoints in deployment/docs/03-API.md
  • Added curl examples and request/response schemas for all new endpoints
  • Categorized endpoints by type (Authentication, Queries, Notebooks, Sessions)

Fixed​

Authentication Preservation:

  • Critical fix: setup_auth no longer erases existing authentication
  • Added check for existing auth before clearing credentials
  • Users can now switch between HTTP and MCP stdio modes without re-authenticating
  • Preserves user experience when switching interfaces

Code Quality:

  • Refactored re_auth to use de_auth internally (DRY principle)
  • Improved separation of concerns in authentication flow
  • Better error handling in HTTP wrapper

Changed​

Version Synchronization:

  • Updated all version references across codebase to 1.3.2
  • Synchronized versions in package.json, src/index.ts, src/http-wrapper.ts, README.md
  • Consistent versioning across all documentation files

[1.3.1] - 2025-01-24​

Added​

MCP Auto-Discovery Tool:

  • New MCP tool auto_discover_notebook for Claude Desktop/Cursor integration
  • Automatically generates notebook metadata via NotebookLM (30 seconds vs 5 minutes)
  • Zero-friction notebook addition: just provide URL, metadata is auto-generated
  • Parity with HTTP API: MCP clients now have same auto-discovery capability

Documentation:

  • Added docs/CHROME_PROFILE_LIMITATION.md documenting Chrome profile conflict
  • Documented current limitation: HTTP and MCP stdio modes cannot run simultaneously
  • Added roadmap for v1.4.0: Separate Chrome profiles by mode

Fixed​

Critical Compatibility Fix:

  • Disabled CompleteRequestSchema handler causing crashes with Claude Desktop
  • Fixed: "Server does not support completions" error on connection
  • Claude Desktop now connects successfully without modifications

Changed​

Tool Documentation:

  • Updated add_notebook tool to recommend auto_discover_notebook first
  • Clarified when to use manual entry vs auto-discovery
  • Added fallback workflow if auto-discovery fails

README Updates:

  • Added warning about HTTP/stdio mode conflict (temporary until v1.4.0)
  • Added Chrome profile limitation to roadmap as priority feature
  • Updated feature descriptions to mention MCP auto-discovery availability

Known Issues​

Chrome Profile Locking:

  • HTTP server and MCP stdio modes cannot run simultaneously
  • Both modes use same Chrome profile, causing "resource busy" errors
  • Workaround: Choose one mode at a time, or stop HTTP daemon before using Claude Desktop
  • Fix planned: v1.4.0 will use separate Chrome profiles automatically

[1.3.0] - 2025-01-23​

Added​

Auto-Discovery Feature:

  • New endpoint POST /notebooks/auto-discover for autonomous resource discovery
  • Automatic metadata generation by querying NotebookLM itself
  • Progressive disclosure pattern inspired by Claude Skills best practices
  • Validation of auto-generated metadata (kebab-case names, description length, tags count)
  • Retry logic for metadata generation (max 2 attempts with 2s delay)
  • New field auto_generated: boolean in Notebook schema
  • Complete documentation in deployment/docs/07-AUTO-DISCOVERY.md

Key Benefits:

  • Autonomous resource discovery: Orchestrators can find relevant documentation without manual intervention
  • Zero-friction notebook addition (30 seconds vs 5 minutes manual setup)
  • Self-organizing documentation library
  • Progressive disclosure pattern optimizes token usage and API rate limits

Changed​

Documentation:

  • Updated API documentation with auto-discovery endpoint details
  • Added progressive disclosure pattern explanation
  • Enhanced README with auto-discovery feature showcase
  • Version bumped to 1.3.0 across all package files

[1.1.2-http] - 2025-01-21​

Added​

HTTP REST API Wrapper:

  • Express.js server exposing the MCP API via HTTP REST
  • 8 documented REST endpoints (see docs/03-API.md)
  • CORS support for n8n/Zapier/Make integration
  • Network configuration via environment variables (HTTP_HOST, HTTP_PORT)
  • Listening on 0.0.0.0 by default for network access
  • Enhanced logs with version, configuration, and available endpoints

Complete Documentation:

PowerShell Automation Scripts:

  • scripts/install.ps1 - Automated installation with checks
  • scripts/start-server.ps1 - Startup with pre-checks
  • scripts/stop-server.ps1 - Clean server shutdown
  • scripts/test-server.ps1 - Validation tests (health, notebooks, ask)

Deployment Package:

  • Isolated and clean deployment/ directory
  • PACKAGE-FILES.txt file listing required files
  • Ready for distribution via Git or npm

Fixed​

Critical Bug - Windows Authentication:

  • Issue: chrome_profile/ remained empty after Google authentication
  • Cause: Windows filesystem does not immediately flush writes
  • Solution: Added a 5-second delay before closing Chrome
  • File: src/auth/auth-manager.ts line 966
  • Impact: Persistent authentication now works on Windows

Bug - Streaming Detection:

  • Issue: Truncated responses or placeholders returned ("Getting the context...")
  • Cause: Stability threshold too low (3 polls) and missing NotebookLM placeholders
  • Solution:
    • Added NotebookLM placeholders ("getting the context", "loading", "please wait")
    • Increased stability threshold to 8 polls (~8 seconds)
  • File: src/utils/page-utils.ts lines 51-53 and 210
  • Impact: Complete and reliable responses (tested up to 5964 characters)

Bug - System Text in Responses:

  • Issue: Each response contained "\n\nEXTREMELY IMPORTANT: Is that ALL you need..."
  • Cause: FOLLOW_UP_REMINDER constant added after text cleanup
  • Solution: Removed the constant and its usage
  • File: src/tools/index.ts lines 30-31 and 791
  • Impact: Clean responses, only NotebookLM content

Changed​

Log Improvements:

  • Added server version in startup banner
  • Display of configuration (Host, Port, network accessibility)
  • List of available endpoints at startup
  • Colored and structured logs via utils/logger.ts
  • Format: log.success(), log.info(), log.warning(), log.error(), log.dim()

Configuration:

  • Documented and standardized environment variables
  • .env support with dotenv (optional)
  • Sane defaults: HTTP_HOST=0.0.0.0, HTTP_PORT=3000, HEADLESS=true

Compatibility:

  • Maintained 100% compatibility with original MCP stdio mode
  • No breaking changes to existing features

[1.1.2] - 2025-01-20​

Added​

  • Support for Claude Code as MCP client
  • Improved documentation for installation

Fixed​

  • Executable permissions for npm binary
  • Reference in package.json

[1.1.0] - 2025-01-15​

Initial version of the original NotebookLM MCP Server project by Please Prompto!

Added​

  • MCP server for NotebookLM via stdio protocol
  • Persistent Google authentication
  • Browser session management with Playwright
  • Multi-notebook support via library
  • Streaming detection with stability
  • Stealth mode anti-detection
  • MCP tools: ask_question, setup_auth, get_health, etc.

Legend of Change Types​

  • Added - New features
  • Changed - Changes to existing features
  • Deprecated - Features soon to be removed
  • Removed - Removed features
  • Fixed - Bug fixes
  • Security - Vulnerability fixes

Notes:

The 1.1.2-http version is a major extension of the original project that adds:

  1. Complete HTTP REST API wrapper
  2. Production-ready deployment package
  3. Comprehensive documentation (5 guides + scripts)
  4. Critical fixes for Windows
  5. Ready for Git/npm publication

All changes respect the original MIT license and maintain compatibility with the original MCP stdio mode.