What's new at Trove.

A calmer home for guides and reading lists

the Living Book guide and How to Build Agents reading list now use Trove’s Reading Room layout, with section navigation and reusable book cards that keep sources close to the text. Existing links still work.

Deleting your account now stops your subscription

your personal plan is cancelled right away, and a team plan you own stops renewing while your teammates keep access until the end of the period they paid for. Time already paid for is not refunded, and Settings now says so.

CandleKeep no longer adds a .candlekeep folder to your repositories

the per-project notebook your agent keeps (which topics it already asked the library about, and your active shelf) now lives in your home directory instead of inside the repo, so it never shows up in git status. Existing folders are cleaned up automatically the next time a session starts there. A .candlekeep folder remains only where you created a project shelf with ck init, because that file is meant to be committed.

Scanned and image-only books are now flagged instead of silently passing

a PDF whose only text was a publisher footer, or an EPUB made of pictures, used to import as "ok" with nothing to read. The importer now recognises text page by page, and when a file still has almost no real text the book is marked Unreadable source in Library Health, in ck items list and ck items read, and to your agents, with the reason and what to upload instead.

Long scanned books no longer time out part-way

the processing limit now scales with the page count, so a large scan gets the time it needs instead of failing at a fixed cutoff.

Your agent consults your library again when you have several manuscripts

the information CandleKeep puts in front of your agent at the start of each session had grown past a size limit in Claude Code, which silently replaced most of it with a short preview. Everything below the cut vanished, including the instruction telling your agent to check your library, so on accounts with a dozen manuscripts it simply stopped doing it. That block is now under half the size, and it can no longer outgrow the limit no matter how many manuscripts you keep.

Manuscripts now tell your agent when to write to them

each manuscript has a new "Log when" line describing the kind of experience that belongs in it, for example "a prod-only bug that took over 30 minutes to find". Your agent reads that line to pick the right manuscript at the end of a session, which matters most when several of your manuscripts cover similar ground. We filled it in from your existing criteria, and you can edit it on the Manuscripts page.

Your agent now checks a per-project notebook before non-trivial tasks

it keeps a small ledger of which topics this project has already asked the library about and what came back, sends the librarian on its own for a topic it has never seen, and when the only matching books are on the marketplace it stops and asks you with a one-keystroke offer instead of sending readers to books you do not own.

New on the blog: Getting an agent to look things up on its own

an engineering account of the new librarian trigger: what ten books on foraging, curiosity, and bandit algorithms said about when to search, the three instruction wordings we tested on real sessions, and why the one that worked leaned on legitimacy rather than force. Read it at /blog/when-should-an-agent-go-look-something-up.

You now get an email when we reply to your support ticket

it carries our reply and a link straight to the conversation, so an answer no longer waits for you to check back on your own.

You can see and undo a book's edit history

ck items history <id> lists the earlier versions of a book and ck items restore <id> --version N puts one back. Every edit has always been saved; now you can actually get one back. Restoring is itself safe: the current text is saved first, so a restore can be undone by another restore.

Our Security and Privacy pages now say exactly where we stand

the Security page describes our own controls plainly (no more "SOC2 roadmap"; the vendor certifications for Clerk, PostHog, and Railway remain), and both pages now state it outright: we don't train models on your content and we don't use it to train anyone else's. The full scope is in the new "No Model Training" section of the Privacy Policy.

Your agent no longer says "your library is full" when it isn't

when adding several books at once hit a momentary write collision, the CLI now waits and retries on its own, and the agent tells a real plan limit apart from a passing hiccup instead of reading every failure as the item cap.

Two agents writing to the same book no longer overwrite each other

editing a book means downloading it, changing it, and uploading it back, so anything another agent added in between used to be silently erased. This hit shared books and LLM wikis hardest, since every session writes to them when it finishes a task. Writes now check whether the book changed while your agent was working: if it did, the write is refused and the agent is told what changed so it can redo the edit on the current text instead of overwriting someone's work. Adding content never conflicts at all.

Adding several books at once no longer drops one silently

when your agent subscribed to a handful of marketplace books or uploaded several files in parallel, the database sometimes rejected one of them, and the agent read that as "your library is full" and quietly left the book out of your reading list. Those collisions are now retried automatically, the CLI waits and retries anything left over, and your agent can tell a real plan limit apart from a passing hiccup — so it only mentions upgrading when your library genuinely is full.

Your Claude Code agents find ck even when your terminal's PATH hasn't caught up yet

right after a fresh install, a Claude Code session launched from the same terminal (or from a GUI app or IDE) could not see the new binary, and the CandleKeep agents wrongly told you the CLI was not installed. The plugin now locates ck in its usual install locations on its own, the session start hook repairs PATH for the whole session, and the installer tells you to open a new terminal before launching Claude Code.

The "How to Build Agents" reading list is live at /how-to-build-agents

every source behind the lecture (the agent company that runs CandleKeep, the Anthropic and OpenAI prompting guides, Anthropic's engineering blog), each linked to the book your agent can read.

The Anthropic Engineering Playbook now carries 24 posts in 17 parts

the full text of Building Effective Agents, Effective Context Engineering, Writing Effective Tools, Advanced Tool Use, Code Execution with MCP, Effective Harnesses, Demystifying Evals, Claude Code Best Practices, and more, one post per page.

Your LLM Wiki is now part of onboarding

right after signing up you see the wiki you already own, learn that it is private to you, and choose how your agent writes to it: Auto (writes on its own, one-line notice) or Ask (proposes, waits for your yes). New accounts start in Ask mode until they choose; wikis that already exist keep their current setting. You can switch at any time on the Manuscripts page. The wiki, like every book in your library, is private to your account unless you share it — and sharing a book with your team now asks you to confirm first.

Full-library MCP calls no longer break on unusual line-break characters

a title carrying an invisible Unicode line separator (common in text pasted from Apple apps or EPUB metadata) used to cut list_items and library_summary off mid-response for Cowork and other Python-based agents. Titles, authors, and descriptions are now cleaned on the way in, every MCP response is guarded on the way out, and those separator characters are removed from existing books automatically.

Hitting the monthly read limit now sends you somewhere real

the upgrade link goes to your billing page (the old /pricing link redirects there too), and your agent gets a clear "stop, don't retry" answer with your usage and reset date instead of hammering the wall with blocked calls.

ck auth whoami shows your monthly reads

Reads (month): 96 / 100, so you can see where you stand before you hit the cap.

ck --version now tells you if you're current

it prints (latest) when you're up to date, or the newer version and how to get it when you're behind. ck version does a live check (--json supported for agents).

ck update — update the CLI on demand

one command brings ck to the latest release, for you or your agent. It works even with auto-update opted out.

Agents get told when your CLI is stuck on an old version

if silent auto-update is disabled or failing and you're behind, regular commands print a short note (at most once a day) telling the agent to run ck update, so it can fix the problem itself mid-session.

Update opt-outs, clarified

CK_NO_AUTO_UPDATE=1 now stops the CLI only from replacing itself; it still checks the latest version once a day so it can tell you you're behind. To stop version checks (and their network call) entirely, set CK_NO_UPDATE_CHECK=1.

ck items import <file.md> — import a whole book in one command

the CLI splits your markdown on chapter boundaries, uploads it chunk by chunk with progress and automatic retry, and gives you a ready book. No more hand-rolled append loops.

Appending to a book is now instant, no matter how big the book is

appends touch only the pages they add instead of rewriting the entire book, so bulk imports are fast and light for books of any size.

Books have a clear 30MB size ceiling

beyond it, the API and CLI tell your agent exactly what to do: split the work into multiple books at meaningful divisions (parts, volumes, sections) and import each one.

Importing very large books could make the service unavailable

a bulk import of a 68MB book caused an outage on August 27; reads and writes of large books are now bounded so one big book can never take the site down.

Project shelves — declare the books your project depends on

run ck init in a repo to create .candlekeep/shelf.json, a checked-in list of the books agents should read for this project. Commit it and teammates inherit the same shelf. ck shelf sync keeps the file and your CandleKeep shelf in step, in both directions.

Agents now see your project shelf automatically

every session starts with a compact index of the shelf's books and topics, so agents in any mode (not just plan mode) go straight to the right book when a task touches its topics. In our tests this took agents from 12 exploratory commands to 2 direct reads.

The Librarian's Note

when a book outside your shelf materially improves a task, the librarian shows one short note at the end of the session: which known author's book helped, the exact changes it shaped, and a one-key choice to add it to the project shelf. Declining is remembered; you're never re-pitched.

The librarian no longer adds books to your library by itself

marketplace books are read freely during a task, but nothing is subscribed or shelved without your explicit yes. Your library stays curated by you.

ck items read now accepts --pages 7-9

the flag agents naturally reach for works alongside the existing "<id>:7-9" form (same for ck items get with a single page).

Team features now require an active team plan

creating team manuscripts, sharing items to a team, and the Pro upgrade that comes with team membership all require a completed team checkout. Teams created without activating the plan no longer grant these.

Edit a single page of a book without re-uploading the whole thing

ck items put <id>:<page> replaces just that page, and ck items get <id>:<page> fetches just that page. Find the page number with ck items toc <id>. This works on books of any size and is now the recommended way for agents to update a wiki's Index and Changelog pages.

Large books were stuck append-only

once a book grew past 1 MB, whole-document ck items put was refused with "413 Payload Too Large", so existing pages (including an LLM Wiki's Index and Changelog) could no longer be edited from the CLI and auto-maintained wikis silently fell behind. Page-scoped writes remove that ceiling; the 413 message now points to them.

Wiki instructions and the book-writer agent now use page-scoped writes

new LLM Wikis are seeded with instructions to append new pages and edit Index/Changelog page-by-page, and to report a failed write instead of silently skipping the bookkeeping.

Choose which coding agents CandleKeep lives in

Settings ▸ Connected agents shows every supported agent with its status on your machines and an on/off switch. Turn one off and CandleKeep removes only its own files from that agent, on every machine you use. Machines still running an older CLI apply the change once they auto-update.

What your machines report

to show statuses per machine, the CLI sends each machine's agents, CLI version, hostname, and OS. Turn off agent-host reporting under Settings ▸ Privacy and hostname and OS are no longer stored.

Same control from the terminal

ck setup --disable <agent> / --enable <agent> apply immediately, and ck setup --status shows what is installed where.

CandleKeep installs itself into agents you add later

install Cursor or Codex after CandleKeep and the CLI notices on its next run and sets it up for you, unless you have switched that agent off.

Your Value report now shows the books your agents are asking for

when an agent searches your library and comes up empty, the librarian notes a real published book that would have answered it. The Value page and ck value now surface those suggestions, grouped by book with the original searches under each, so you can judge the suggestion and go get the book yourself. Suggestions are AI-matched at search time, so a one-line note reminds you to verify before you buy; the section stays private and never appears on your public profile.

Editing a published book no longer sends it back to review

replacing content, appending pages, renaming, re-uploading a file, or reprocessing a book that is listed on the marketplace now takes effect immediately. The listing stays live and visible to new subscribers; the "Listing returned to review" warning and notification are gone. New listings still go through the initial marketplace review.

Your new Value report shows what your library actually does for you

the Value page on the dashboard tells the story session by session: what you were researching, which books your agents read and how deep they went (full content or just the table of contents), how often your library had something relevant, plus a reading heatmap, your most-read books, and a few fun comparisons ("your agents read the equivalent of War and Peace this month"). Looking at the report never counts as a read.

ck value brings the same report to your terminal

same numbers as the web page, rendered for your terminal. Pick the window with --since (--since 7d or --since 2026-07-01, default 30 days), show more sessions with --sessions N or --all, and get --format markdown or --format json for piping into other tools.

The Value report's wiki section now shows your most-read wiki entries

see which LLM Wiki entries your agents come back to most, with the entry content readable right there: expand an entry on the web report to read it in place, and ck value lists the top entries with their titles and a short snippet.

Share your reading stats with a public profile

publish a profile at getcandlekeep.com/u/your-handle from the Value page or with ck value share. It shows your reading heatmap, totals, and fun stats — book titles stay hidden unless you reveal them, your research questions are never public, and you can revoke the page anytime (your handle stays reserved, so republishing brings back the same URL).

The Pro Library section is readable in light mode again

its headings and links used colors designed for a dark backdrop, making the Pro pitch on the marketplace nearly invisible for light-theme visitors. The section now uses theme-aware colors that stay legible in both modes.

Marketplace badges no longer break in dark mode

the "in your library" state, "Free This Week" chips, listing badges, and the Pro filter pill's hover state were styled for light mode only and rendered as glaring or unreadable elements for dark-mode visitors. They now adapt to your theme.

Landing-page label pills have a visible outline in light mode

two labels on the homepage drew their border with a faint white ring that was invisible on a white background; they now use a real border that shows in both themes.

The press page and LLM Wiki page no longer ask visitors to sign in

both public pages were accidentally behind the sign-in wall, so anyone following a link from the homepage or a search result was bounced to a login screen instead of the content.

The theme toggle now controls every part of the page

if your CandleKeep theme differed from your operating-system theme, some elements followed the OS instead of your choice and rendered with the wrong colors. Everything now follows the in-app toggle.

You can now hide the CandleKeep citation box

the boxed "Read / Learned / Worth remembering" summary at the end of responses is now a setting. Turn it off in Settings → Citation block or with ck config set skill.citationBlock off, and it stays off across every agent you use CandleKeep in — Claude Code, Cursor, Windsurf, GitHub Copilot, Gemini CLI, Roo, Codex, and Claude Cowork. Inline citations like (Book Title, p. 12) still appear, so answers stay sourced. The box stays on by default.

Reads through the Cowork plugin now count toward your monthly reads

books opened from Claude Cowork were never being counted, so they didn't appear in your reading stats and didn't draw down the Personal monthly read allowance. They now count the same as reads from the CLI or the Claude Code plugin. If you read mainly through Cowork on a Personal account, you'll now see that usage on your billing page and can reach the monthly cap where previously you couldn't.

Cowork reads now show up in your reading history

because those reads weren't recorded, they were missing from your reading map, your "changes since you last read" notices, and (for authors) the reader-channel breakdown in writer analytics. Cowork is now its own channel in all three, shown separately from CLI and plugin reads.

A privacy setting for which coding agent you use

CandleKeep now records the name of the agent running it (Claude Code, Codex, or a plain terminal) so we can see which agents people actually use and put our effort there. It's the agent's name only — never your prompts, files, or commands. There's a new Privacy section in Settings to switch it off, which takes effect immediately without updating your CLI.

Staff access to your data is now logged, and you can read the policy

when a CandleKeep admin opens a customer's library, sessions or support tickets, that access is recorded in an append-only audit trail: who, what, when, and from where. Reads are logged, not just changes. The new Administrative Access section on the Security page spells out exactly what is captured.

Manuscripts are counted one way for everyone

your LLM wiki used to sit outside your plan's manuscript allowance, so Personal accounts quietly had room for two. Now it's simply a manuscript like any other: Personal includes one, Professional is unlimited. If you're at your limit, the message now tells you which book is holding your slot and that switching it Off frees it up — instead of just refusing.

Your research questions no longer live forever

the question you type when your agent reads from your library is now automatically erased after 90 days. Your reading history and library stats are untouched; only the question text goes.

The startup script CandleKeep runs in Claude Code now ships inside the ck command itself

previously your machine downloaded that script from our servers and ran it on every session start. It is now built into the CLI you installed, and the plugin files we do still send are checked before anything is written to your disk. Once your CLI has auto-updated, CandleKeep behaves exactly as before; changes to the startup script now arrive with the CLI's normal automatic update.

Turning off a manuscript's auto-update now actually quiets it

unchecking "Auto-update" used to make your agent louder, not quieter: instead of writing silently it started asking for your approval at the end of every session, and the only real off switch was a separate "Pause" button on the manuscript card. There is now one clear choice per book — Auto (writes on its own, one-line notice), Ask (proposes and waits for you), or Off (your agent won't touch or even mention the book) — and all three are in the same place, on both the card and the edit dialog. Each option spells out exactly what your agent will do.

ck auth login now only accepts the sign-in it started

the CLI briefly opens a local port while you authorize in the browser. It now generates a one-time secret for each login and refuses any hand-off that doesn't carry it, so a random web page you happen to have open can't push its own key into your terminal and quietly point your library at someone else's account. The local port is also loopback-only now, and it closes as soon as the sign-in finishes or gives up (15 minutes) instead of staying open forever. Update your CLI (ck self-updates on the next command) to get the fix.

Books that didn't import fully now say so — and you can fix them in one click

a book could land in your library looking fine (normal page count, opens normally) while most of its pages were actually blank, and nothing told you. Your library now spots these and flags them as "partly imported", both on the book's page and in Library Health. A new Reprocess button re-imports the book from the file you originally uploaded; the current version stays readable the whole time it rebuilds, so nothing is lost if it fails.

Reprocess works from the terminal and from your agent too

ck items reprocess <id> re-imports a book, and ck items list now shows the "partial" warning with the exact command to fix it. Coding agents can read that and offer to run it for you. Requires the latest CLI (ck setup to update).

CandleKeep installs into Claude Cowork and Desktop as a single plugin

no more pasting a connector URL and then separately uploading a skill zip. Open Customize → Plugins → Add marketplace, paste CandleKeepAgents/candlekeep-marketplace, and click Install on CandleKeep. That one step adds the research workflow, the four specialist agents, and the connection to your library together, and clicking Update on the marketplace keeps it current — no more re-downloading a zip to get the latest behavior. (We're under review for Anthropic's official plugin catalog; once listed, you'll be able to find CandleKeep straight from Browse plugins.)

Cowork now runs the same four-agent research pipeline as Claude Code

a librarian finds the right books across your library and the marketplace, a reader pulls the exact pages and cites them, and separate agents handle writing documents and cleaning up book metadata. Answers come back with the same page-level citations you get in the terminal.

Professional now uploads files up to 500MB

Personal keeps its 100MB limit, and Professional gets 5× the headroom so large scanned or image-heavy PDFs and long manuscripts go straight in. If a file is over your limit, the upload dialog now shows exactly what you'd unlock by upgrading instead of a generic error.

Signing up on your phone is smoother

CandleKeep runs in your terminal, so mobile signups now tell you that up front and get you to the handoff faster: pick up where you left off on your computer with a single command that signs you in, installs the CLI, and connects your agent. The emailed and QR/copied links now drop you straight onto a page with that one command — no extra sign-in or wizard first. And if you email yourself the link but don't finish, we'll send a short, friendly reminder or two with the command, then leave you be.

Agents flag broken books on their own

when the librarian or reader agents touch your library for a task and notice a book that failed to process or came through empty, they now mention it in their answer instead of waiting for you to ask them to "verify" your library.

CandleKeep is calmer by default in Codex

on the default setting, Codex no longer reaches for your library on every single task. It now consults CandleKeep when a task genuinely benefits (code review, security, design/UI-UX, architecture, agent development, research) and skips trivial or mechanical work like typo fixes, renames, and git commands. If you want the old always-on behavior or something quieter, the Smart, Ask, and Off options in Settings still control it — update your CLI (or run ck setup) and the change applies on your next Codex session.

Your library now tells you when an upload goes wrong

if a book fails to process, comes through empty, or stalls, you get a notification right away instead of finding out later. Failed books now explain what actually happened and what to do next (e.g. "this looks like a scanned PDF without OCR") instead of a generic "processing failed."

New Library Health view

a single place to see everything that needs attention: failed uploads, empty books, stalled processing, books awaiting confirmation, and suspected duplicates, each with a clear next step.

Confirm we got the right book

when we're not fully sure we identified an uploaded book correctly, we now ask you to confirm or fix its title and author, so a wrong auto-detected title can't slip into your library unnoticed.

Uploads now catch duplicates before they pile up

when you upload a book that looks like one already in your library, CandleKeep asks what you want to do instead of silently adding another copy: Replace the old one, Add anyway (for a different edition or translation), or Cancel. It matches on similar titles, so re-uploading the same file — or retrying a failed upload — no longer leaves you with three or four copies of the same book. Works in the upload dialog and in the CLI (ck items add, with --replace and --force flags for scripts).

Share a book you own with your team from the CLI — no admin role needed

ck team access grant <id> used to refuse anyone who wasn't a team owner or admin, even for a book they own, telling them to become an admin. Now any team member can share their own book with the whole team (read-only), matching the website's "Share to team" button. Scoping a share to a specific department or granting write access still require an admin.

Reader Requests — see exactly what your readers wanted your book to cover

when an agent reads one of your books to answer a question but a specific sub-topic is missing, that gap now goes straight to you, the author. Head to Writer Analytics to find your Reader Requests, right next to your read stats: a summary across all your books on the overview, and a per-book panel on each book's analytics page showing the missing topics ranked by how many readers asked (or run ck items gaps from the CLI). Extend your book to cover a topic, mark it covered, and every reader who asked gets notified that the book they read now answers their question.

Your reads now help improve the books you rely on

when your agent reads a relevant book that turns out to be missing the specific thing you needed, it quietly lets that book's author know (as an abstracted topic, never your actual question or any personal details). Authors use these signals to decide what to write next, so the books in the marketplace keep growing toward what readers actually ask for.

Billing portal is more resilient for former subscribers

if you previously had a personal subscription and later got Pro through a team, an old canceled subscription record could cause "Manage Subscription" to fail to open the billing portal. The portal now always uses your active subscription and ignores stale records.

Team members can now read documents shared with their team

team-shared books showed up in your library list but failed to open with an "Items not found" error when read from the CLI or an agent. Reading now works consistently with what the list shows, as long as your team's subscription is active.

Your activation setting now works on Codex

changing the plan-mode activation in Settings (Smart, Ask, or Off) previously only affected Claude Code; Codex kept checking your library on every task regardless. Now Codex respects the setting too: Off removes the auto-check, Smart checks only when the task overlaps your library, and Ask prompts first. Update your CLI (or run ck setup) and the change applies on your next Codex session.

Manuscript cards now show when the book was last refreshed

the timestamp on each manuscript card used to reflect the last time you edited the manuscript's settings, which could disagree with the "Updated" time inside the book itself. Cards now read "Refreshed <time>" based on when the book content actually changed (including your agent's automatic updates), so the list and the book view always agree. A manuscript with no book yet shows "Created <time>".

You'll now hear about your monthly reading limit

when you reach your monthly free-read cap, your agent now clearly tells you and shows how to upgrade for unlimited reads, instead of quietly failing to fetch content. Reading resumes when you upgrade or at the start of the next month.

Change your account email in Settings

head to Settings → Account to update the email on your CandleKeep account. Add a new address, verify it, and set it as your primary — your CLI and API keys keep working unchanged. Password and security settings live there too.

Choose which emails you get

a new email preferences page at /settings/email lets you keep account & billing emails (always on) while muting product updates, onboarding tips, or research invites independently. Every non-essential email now has a one-click unsubscribe that mutes just that kind, not everything.

Report bugs and request features straight from the CLI

run ck report --title "..." --body "..." to file a bug report or feature request without leaving your terminal. Pass a large diagnostic dump with --body-file <path> or pipe it in with --body -, and set --type bug|feature|idea if you like. Your report shows up on your Support page where you can follow the conversation, and your Claude Code agent can file one for you when it hits a wall.

We'll tell you when a book you asked for arrives

when your agent searches for something your library can't answer, CandleKeep now remembers that you asked. If we later add a book that fills that gap, you'll get a heads-up right in your terminal the next time you run ck, with a one-line command to add the new book to your library.

Compare CandleKeep to other tools — new comparison page covering Obsidian, NotebookLM, Notion, Glean, and building your own RAG at /compare

Manage your team from the CLI

owners and admins can now run the whole team admin flow in the terminal with ck team: create departments, invite people, add them to departments, and grant or scope documents (team-wide or to a single department) with read or write access. Run ck team access matrix to see who can read what. Your Claude Code agent can do it conversationally too — "add the onboarding handbook to the Engineering department" just works.

Clearer no-terminal setup for Claude Desktop & Cowork

the install page now leads with the no-terminal path so chat-based agents and Desktop/Cowork users aren't sent to a terminal command that doesn't apply to them, and the Cowork guide has a plain-text "Copy-paste version" you (or Claude) can follow without screenshots. Added troubleshooting for when behaviour seems stale after an update — re-download the skill bundle to refresh it.

Per-seat Team plan

the Team plan is now $30 per seat, and you buy exactly the number of seats you need: invite a co-founder for 2 seats or roll out to a 50-person org, you only pay for the people on the team. Add or remove seats anytime, and manage billing from the team page.

Shared team manuscripts

your whole team's agents can now read from a shared, living knowledge base. When one engineer finishes a refactor and their agent writes up what changed, that knowledge lands in the team's shared manuscript — so the next teammate's agent starts from what the team already learned instead of from scratch.

Team admin panel with department-based access

owners and admins get a dedicated panel to organize members into departments and grant specific books to specific departments. Share a book with the whole team, or scope it to just Engineering or just Sales, and each member's agent only ever sees the books they're cleared for. An access matrix shows at a glance which departments can read which documents.

Connect MCP clients with your API key

the CandleKeep MCP server now accepts your ck_ API key as a Bearer token, so you can wire up any MCP client directly without the OAuth connector flow. Grab a key on the Developers page, where there's now a copy-paste MCP setup section.

The team page now shows live seat usage so you always know how many seats are in use versus purchased before you invite someone new.

LLM Wiki auto-updates now actually write — Claude Code was silently blocking the wiki write in auto mode; ck setup (and the background updater) now grant the needed permissions

Agents can now locate your wiki manuscript reliably — the session context includes the manuscript ID and ck ms show accepts the book ID as a fallback

The Windows install one-liner (irm https://getcandlekeep.com/install.ps1 | iex) now works in Windows PowerShell — it previously crashed immediately before installing anything

CLI downloads no longer fail on slower connections — the installer and auto-updater now allow slow-but-progressing downloads to finish, retry transient network errors automatically, and show clearer error messages when a download truly can't complete

Share your library

make any book public from its ⋯ menu, then share your reading list. Your public profile page at getcandlekeep.com/share/you now shows the books you've added to your showcase, and the share dialog can copy your list as a clean markdown table to paste into a gist, README, or post. Books are private by default — only the ones you add to your showcase ever appear.

LLM Wiki

your library now includes an auto-maintained "My LLM Wiki" book that your coding agent updates on its own after each session, capturing the concepts, decisions, and patterns you work through so your knowledge compounds over time. No approval prompts — the agent writes it and posts a one-line note, and every change is reversible from version history.

Auto-update for any manuscript

switch on "Auto-update" for a manuscript (in the dialog, or with ck ms create/update --auto-update) and your agent maintains it automatically at the end of each session instead of asking you to approve every addition.

Per-manuscript writing instructions

give any manuscript its own instructions for how the agent should write and structure it — tone, layout, what to include, how to link pages — and the agent reads and follows them. View them anytime with ck ms show <id>.

Control when CandleKeep activates in plan mode

a new Settings page lets you choose how the skill triggers while you plan: Always (check your library every time), Smart (only when your library likely covers the task), Ask (confirm first), or Off. Set it in Settings or from the terminal with ck config set skill.planMode <mode>, and it applies to your next session — no reinstall needed. Already-activated power users can dial it down; the default stays Always.

Books shared with you or subscribed from the marketplace no longer show the original owner's private shelf names. Your library and ck items list now only ever display shelves you created yourself.

Subscribed books no longer surface the owner's private team name in your dashboard.

New tier names: Personal and Professional

Free is now called "Personal" and Pro is now called "Professional" across the app, pricing page, and CLI. Same product, clearer identity.

New pricing for new signups

Personal is $0/month with 5 items and 100 agent reads per month, plus 1 manuscript. Professional is $19/month with everything you had on Pro: unlimited reads, 200 items, marketplace publishing, unlimited manuscripts.

Existing users keep their plan

if you signed up before today, your limits and price are locked. Free users still have 20 items and 500 reads per month. Pro subscribers stay on $10/month. Nothing changes for you.

The trial-expired modal and billing page now show the right limits for your account — old accounts see their original caps, new accounts see the new Personal caps.

One install hub for every surface

visit /install to pick how you'll connect: macOS/Linux one-liner, native Windows PowerShell, no-terminal Cowork/Desktop, or browser-only. The page auto-detects your OS (and defaults mobile visitors to the browser tab) so you see the right command first, but all four paths are visible side-by-side.

Help drawer in the dashboard

click the new help button in the dashboard header, or press ? anywhere, to open a quick panel with install links for other surfaces (including a copy-pasteable prompt you can send straight to your AI agent), common tasks (upload, invite, marketplace, manuscripts), and resources (Slack, support, what's new).

CandleKeep in Claude Cowork & Desktop

paste one URL, upload one zip, and Claude can search your library, read pages with citations, and draft knowledge docs. No terminal needed. Walkthrough at /install?surface=cowork.

Promote any book to a manuscript

every book in your library now has a "Promote to Manuscript" button on its detail page. Turn an existing book into a living manuscript without creating a duplicate. Same on the CLI: ck ms create "Title" --item-id <book-id>. Reattaching to a book that has an archived manuscript on it just works — the archived one quietly steps aside, keeping its topics and criteria intact.

Codex CLI integration

if you use OpenAI Codex CLI, ck setup now provisions a Codex-tuned skill plus four parallel subagents (librarian, reader, writer, enricher) optimized for GPT-5.5. Codex can spawn library research as a fan-out — librarian picks the books, multiple readers extract findings in parallel, you get a cited synthesis.

New /marketplace/activity page shows recent book additions and updates across the marketplace.

Screenshots in support tickets

attach images to new tickets and replies, or paste a screenshot directly into the message. Up to 4 images per message, PNG/JPEG/WebP/GIF, 10MB each.

Windows and Cowork are first-class install paths now

both used to live as footnote links on the landing page; they're equal options in the install hub, the onboarding wizard, and the new help drawer. The "Getting Started" sidebar entry was renamed to "Install Guide" and points at the new hub.

/install and /help pages now show the site nav

the install hub and the Cowork/Windows guides used to look stranded; they now have the same header as the rest of the marketing site, so the logo, marketplace, blog, and dashboard are always one click away.

Cowork install screenshots zoom on click

the four Customize-panel screenshots in the Cowork install guide (and the preview on /install) open at near-fullscreen so you can actually read the labels. Esc or click the backdrop to close.

Smarter agent-install prompt

the copy-pasteable "send this to your AI agent" line on /install and in the dashboard help drawer now points the agent straight at https://getcandlekeep.com/install, so it can read the install steps for your environment and run them, instead of guessing what "set up my library" means.

Large books can be edited again

books that grew past the request body limit can now be updated using the new ck items append command, which sends only the new chapter instead of the entire book. Use ck items append <id> --file chapter.md to add at the end, or --prepend to add at the start. Existing ck items put still works for smaller documents.

Books in your library now tell your agent when they've been updated — your agent will note what changed (new chapter, removed section, refreshed content) before answering from the book.

OpenAI Prompting Best Practices book updated

new Chapter 12 on Codex CLI on-disk conventions (skills, subagent TOMLs, AGENTS.md), plus surgical updates to Ch. 3 (GPT-5.5 availability constraint), Ch. 5 (Personality variants), Ch. 7.9 (SDK vs CLI Skills), and Ch. 11 (Codex CLI subagents).

CandleKeep activates proactively in Codex

for code review, UI/UX critique, design feedback, security audits, architecture decisions, and "find improvements" / "research X" queries, the librarian fires as Codex's first action and grounds the answer in your library rather than going to the web. Matches the Claude Code behavior you're already used to.

ck no longer crashes when run from inside Codex's read-only sandbox

previously you would see a Rust panic ("Attempted to create a NULL object") on the first ck items list from a Codex session. The CLI now uses a pure-Rust TLS stack and fails gracefully if network is blocked.

Upgraders no longer see a duplicate candlekeep skill in Codex's /skills list

the deprecated install path (~/.codex/skills/candlekeep/) from earlier CLI versions is now cleaned up when you re-run ck setup.

Activity graph tooltip on top row

hovering Sunday cells no longer cuts off the tooltip.

Book suggestions on library miss

when your agent searches for a topic not covered by your library, it now suggests a real book that would help and remembers what it suggested so it won't repeat itself

Sharing on social and search just got better

links to blog posts, marketplace listings, and the home page now render with proper preview cards on Twitter/LinkedIn/Slack, and search engines see structured data (Article, Product, Organization) so results show richer snippets

Cleaner page titles in browser tabs and search

tightened the home title and meta description so they fit Google's display caps and lead with what CandleKeep does for you, not what it is

Writer Analytics

marketplace authors can now see how their published books perform: weekly read rate, unique readers, chapter-level heatmap showing which sections agents reference most, and channel breakdown (CLI vs Plugin vs Web). Privacy-first design hides granular breakdowns until 5+ unique readers.

More reliable ingestion for long books

large, formatting-heavy PDFs and EPUBs (1000-page references, technical AI/RAG books, finance manuals) that previously timed out during processing now ingest to completion. Markdown structure is preserved — headings, tables, lists, code blocks, and italics survive intact instead of being lost to a plain-text fallback.

Resumable ingestion

if a worker restarts mid-upload (during a deploy or transient infra issue), your book picks up from the pages already extracted instead of starting over.

Manuscripts

define knowledge areas and your coding agent automatically proposes book additions at the end of sessions based on what it learned

Manuscripts feature discovery across the app — contextual cards, sidebar badges, and educational content to help you get started

Archived manuscripts no longer count toward your billing quota

PRO pricing copy now accurately reflects available features — removed references to unimplemented revenue sharing and team features

CLI now supports headless login via ck auth login --with-token for SSH, CI, Docker, and Codespaces environments

Marketplace search bar is now always visible and larger — no longer hidden until you enter the All Books view

CLI setup and webapp onboarding now showcase the full marketplace catalog so you can discover relevant books immediately

One-click email unsubscribe with CAN-SPAM/GDPR compliance — native support in Gmail and Apple Mail

Trial transparency — persistent banner showing days remaining, expiration countdown modal, and clear upgrade path

Onboarding redesigned around automatic activation — CandleKeep now fires during plan mode without manual setup steps

Free tier read limit errors now include a direct link to upgrade

Table of contents no longer shows expand buttons for chapters without sub-sections

Skill cache now refreshes immediately after admin updates

Librarian agent

automatically finds relevant books from your library and the marketplace during plan mode, no manual browsing needed

Research pipeline now follows a focused librarian-then-reader flow for more precise results

Plan mode integration streamlined to a single research phase

CandleKeep activates reliably in plan mode via improved skill routing

Team plan users are no longer incorrectly downgraded when their trial expires

Team subscriptions now recognized across all billing checks and feature gates

Windows CLI support

CandleKeep CLI now runs natively on Windows via PowerShell, no WSL required

Citation block

agents now cite exactly what they read and how it influenced their answer, with a "Worth remembering" quote for gradual learning

Structured citation block appears at the end of research-informed responses

Research costs reduced by 80% with no loss in accuracy (Sonnet replaces Opus for reading)

Marketplace redesign

magazine-style layout with visual book covers, category-based color palettes, and editorial featured book sections

Book table of contents now shows page ranges for each chapter and section (e.g., "Chapter 1 (p. 5–11)")

Marketplace shows "In Library" badge for books you already own

New users can now sign up directly from the CLI authentication page

CandleKeep now activates automatically during planning and architecture tasks

Your library and marketplace catalog are injected at the start of every Claude Code session

Research sessions are now tracked for better context continuity

Create, edit, and manage markdown documents alongside PDFs

New Book Writer agent creates and edits documents through natural conversation

Library view now shows markdown documents alongside PDFs

Books with missing metadata are now automatically enriched with titles, authors, and descriptions

Table of contents extraction for easier navigation within books

Library view now shows enrichment status and a queue of items needing review

CandleKeep Cloud launches with document library, AI-powered research, and CLI access

Search your library and get cited answers from your books

Upload PDFs and manage your knowledge library from the command line