Usage
Use the Trove dashboard in Claude Code
Updated 15 Nov 2026
On this page
For agents
Load this page when: you need to explain, check, open or turn off the Trove dashboard in Claude Code
- The dashboard is a panel in Claude Code 2.1.287 or later. On a new install it opens by itself in a wide terminal; /trove-panel closes or opens it. You cannot run that command for them.
- To check it, run
trove setup --status --jsonand read thedashboardobject. Ifstateisneeds_update, tell the person to runclaude update. - It sends only usage data: small counts, a few lines of counts for each task and a one-way session key, never prompts, book text or file contents. The person can turn it off with the Usage data switch in the panel's Settings, the Share which coding agent you use switch in Trove Settings, or TROVE_NO_CLIENT_TELEMETRY=1. Never turn a privacy setting off for the person.
- To turn the dashboard off (the panel, book reading while the person works, the notes, the hint and the Pro book line under the reply), run
trove mod settings set isOff true. Trove saves it in the person's account, for all their computers, and it applies from the next Claude Code session. To close only the panel and keep the book reading, the person types /trove-panel. You cannot run it.trove mod settings set isOff falseturns the dashboard on again. - The panel also shows a card that offers to write a book from the person's sessions (Write it, Not now, Never). Once per computer it first shows a list of up to five subjects, each with Write it, and a Close button. The person chooses with the buttons or types
/trove-panel book write|later|never|retry|close, or/trove-panel book write <n>for a row of the list. You cannot run these. - Run
trove setup --disable claude-codeonly when the person asks to remove Trove from Claude Code. It removes all Trove files, the skill included.
The dashboard is a panel in Claude Code. It has one switch for reading: Auto or Off. In Auto, Trove reads the library books that fit your work, and gives your agent short notes with the book and the page. It works for a question, a piece of writing and code.
Before you start
- Trove is installed and
trove setuphas run (see Install). - Claude Code is version 2.1.287 or later. Older versions ignore the dashboard.
- You use Claude Code in a terminal. The dashboard does not draw in other surfaces.
Steps
- Check that it is on. Ask your agent to run
trove setup --status. Theclaude-codeline saysdashboard: on, or it names the Claude Code version that you must update to. - Open the panel. On a new install the panel opens by itself in a wide terminal.
/trove-panelcloses or opens it. - Give your agent a request. In Auto, each message of two or more words goes to a small model, the librarian. A message of one word goes to no model. A reply that only agrees or thanks, such as "ok thanks" or "go ahead", also goes to no model. For each message, the librarian first decides if the message needs books. A quick fact, a command, a change to the format of a text, or chat usually does not need books. For such a message, Trove reads no book, adds no marketplace book and sends no line that names books. When the message needs books, the librarian finds the books in your library that fit the message. For each message, the librarian also decides if the message starts a new task or continues the current task. To decide, it reads your last messages, the last reply of your agent, one short line for each of the newest tool calls of your agent in the current task (at most 30), the books that Trove read in this session with the message that each book was read for, the books that your agent read in this session, and the books that Trove and your agent read in earlier sessions of this project. Trove tells your agent that it reads books for the message. Then Trove starts one reader for each book that fits, and all readers run at the same time. Each book whose pages a reader reads counts toward the monthly reads of your Trove plan, as one read per hour however many pages or parts it reads in that hour. Looking at a table of contents does not. In one task, Trove does not read a book two times. In one session, Trove does not read again a book that it already gave notes from, or a book that your agent read itself.
- Let the readers work. A reader reads the table of contents and picks pages. It reads them and decides if it needs more pages. It stops when it has what it needs or when no page that it can read is left. There is no time limit and no fixed page count. A reader also stops when you set reading to
Offor when the monthly reads of your plan are used up. A new task does not stop the readers of the task before it. - Read the notes. A reader returns notes. Each note is one sentence with the page it comes from. A reader also reads the short lines of the tool calls. It gives a note only when the note can change the work. It leaves out what your agent already has or did, and a note about a step that the work does not do. If nothing on the pages can change the work, the reader gives no note. Trove drops a note when it is empty, when it repeats a note that Trove already has, when its page is not a whole page number that the reader read, or when you set reading to
Offwhile the reader made the note. Trove cuts a note that is longer than 1000 characters at a word end. Apart from the fit check in step 6, Trove does not drop a note because of what it says. There is no limit to the number of notes in a session. - See what your agent gets. Trove sends each note to your agent as hidden context. If your agent is working, the notes go with its next tool call. If your agent already answered, the notes wait. When you send your next request, the librarian checks the waiting notes against it (the 20 newest in one check; older ones go as before). The notes that fit go to your agent with its next tool call. Trove drops the other notes and does not send them. If the check cannot run, Trove sends all the waiting notes. A message of one word, or a reply that only agrees or thanks, such as "ok thanks", takes the waiting notes with it unchecked, marked as notes for your earlier messages. A late note for an older task waits for the next check. If your agent answers without a tool call, the notes that fit wait for your next message. Each note gives the book, the page and the sentence. A label says that the notes come from your library and that your agent must weigh them as advice. The notes do not show in the chat. The panel shows what your books changed in this session.
- Look at the hint. The librarian decides where a task starts, so one task can have many messages. When your agent says that a book changed what it did, one line under the reply names the book and what your agent was about to do, with a
Trove panelbutton that opens the panel. The panel then shows a card with what your agent was about to do, what would have gone wrong, and what it does now. If a marketplace book was added, a card shows it, and you selectKeeporRemove. You can turn off the card animation withMagic scrollinSettings. A Pro book line can also show under the reply, only when the panel itself read a Pro book (never for a read by your agent). On Personal, when the panel could read only the preview of a Pro book in your library, the line names the book and has aGet Probutton. It shows at most once for each book in a session. During a trial, when the panel read a Pro book in full, the line names the book and the day the trial ends, with aKeep Probutton. It shows at most once a day. The button opens the Trove billing page in your browser. Paid Pro sees no line, and the end card takes its place when both would show. - Turn it off. Open
Settingsin the panel and setRead books while you worktoOff. Trove then reads no books for your work. You can set it toAutoagain from the panel at any time.
Prompt to paste
Check that the Trove dashboard is on for Claude Code. Tell me the result and, if it needs an update, the command to run.What you will see
The agent runs trove setup --status and reports the dashboard line:
dashboard.state | What it means |
|---|---|
on | The dashboard files are installed and Claude Code is new enough. |
needs_update | Claude Code is older than 2.1.287. Run claude update. |
claude_not_found | The claude program was not found. Install or update Claude Code. |
unknown_version | Trove could not read the Claude Code version. The dashboard works when the version is 2.1.287 or later. |
not_installed | The dashboard files are not there yet. Run trove setup. |
What the dashboard sees and sends
- The dashboard reads your messages, the replies of your agent and the tool calls of your agent, with their results. It uses this text to find the books that fit your task and to decide where a task starts. For each tool call of your agent in the current task, it also keeps one short line: the tool and its target, such as a file name or a command. Trove removes the secrets that it can recognise from the line. It makes no line for a command that shows a secret, for a call that you refused, or for a private file, such as a
.envfile, a key, or a file outside your project folder. In a command, a path outside your project folder shows only its last name. The librarian and the readers read these lines. The lines are kept in memory only. - This text stays in your Claude Code. The librarian and the readers ask Claude through your own Claude Code, with your login. Trove's server gets your usual library calls, your settings, the book reads and the counts, never this text. The text is kept in memory only, not written to disk.
- In Auto, it gives your agent short notes with the book and page, without asking. The text is marked as advice from a book, not as an instruction.
- With
Marketplace books: On(the default), Trove can add a marketplace book to a free slot in your library, at most one book for each task. The book then gets a reader like any other book. At the end you selectKeeporRemove. If you selectRemove, Trove does not add that book again. You can add it again from its marketplace page while the book is listed. This setting stays on this machine. - When the panel reads a Pro book, it checks your plan with
trove auth whoamiand keeps on this computer only the day, the session key and the ids of the books it showed a Pro line for. It sends no event for the Pro line. It opens your browser only when you press the button of that line.
How Trove knows that a note helped
The notes ask your agent to end its reply with a hidden block. The block names the notes it used and the notes it did not use. For a note that changed its work, the block also has one line with what the agent was about to do, what that would have caused and what it did instead. The panel removes the block from the reply that you see, also while it is written. A changed line becomes a save: the panel shows it as a card, and one line under the reply says that a book changed the work.
The text of the changed lines stays in your Claude Code session, for this session only. Trove's server gets only the numbers. It never gets the text of the notes, of the changed lines or of your agent's replies.
Books from your sessions
The panel can also offer to write a book. When a subject keeps coming back in your finished Claude Code sessions and no book in your library covers it, one card appears at the top of the panel: "A book you could write". It names the subject and says how many sessions it comes from. While the card waits, the panel's top line says A book you could write.
The card has three buttons:
Write itstarts the book. The card then says "Writing your book" with the chapter it is on. When the book is ready, the card says "Your book is ready" and has anOpen itlink. The book is in your library.Not nowhides the subject for 14 days.Neverhides the subject for good.
If the writing fails, the card says why and has a Try again button. After the list, you see at most one card in a session. Only the Off switch of the dashboard stops this feature. Reading set to Off, or Usage data set to off, does not stop it.
The first time, once per computer, the panel instead shows a list called "Books you could write". It appears after Trove has read enough of your finished sessions, and it lists up to five subjects that have no book yet. Each row has a Write it button. Write it on a row turns that row into the card above and starts the book. Close hides the list. The list does not come back, and you see at most one card in a session.
You can use the keyboard instead of the buttons. Type /trove-panel book write, /trove-panel book later, /trove-panel book never or /trove-panel book retry. While the list shows, type /trove-panel book write 2 to write the second subject, or /trove-panel book close to hide the list.
See Log what you learn for the steps and for what stays on your computer.
Usage data
The panel's Settings has a Usage data switch. It is on by default. It says:
Trove collects usage data, linked to your account, to improve the product. No prompts, book text or file contents are sent. You can turn this off.
The data is small counts:
- A few lines for each task. A task finishes when the librarian decides that a new message starts new work, or when the session ends. The panel sends a line when the work goes quiet after a reply, a line when the task finishes, and a new line if late notes arrive. Each line has the number of the task in the session, how many books the panel read, how many books your agent read itself, how many notes the panel sent, how many of these notes your agent used by its own report, and if the hint under the reply showed (a book changed what your agent did). A line also lists each note the panel sent, with the book id, the page and whether your agent used the note. The lines have no titles and no text. Trove keeps one row for each task and one row for each note. If the panel cannot load your library, it also sends one error count with a short word.
- The kind of each book. The first time the panel reads a book that has no kind yet, it makes one small model call in your own Claude Code. The call sees the title, the description and the contents list. It decides whether the book is your own material and in which field of work it is. Trove gets only the book id and these two labels. The first answer stays, and a label never shows to you or to other people.
- A session key. Each call that the CLI makes in an agent session has a session key. The CLI makes the key on your computer. It is the first 16 characters of a one-way hash (SHA-256) of the session ID of your agent. The session ID stays on your computer. The server hashes the key again with your account. Nobody can link the same session across two accounts. The server also stores the key on the records of your calls. Trove uses the key to count the agent sessions that use Trove. It also uses the key to put the reads of one session together.
- A panel mark. A call that the panel makes, not your agent, has a mark that says so. Trove uses the mark to show the reads of the panel apart from the reads of your agent. A read by the panel is a read like all other reads. It counts toward the monthly reads of your plan.
Each of these three settings stops the counts and the session key:
Usage datain the panel'sSettings. It applies to your account, for all your agents, on all your computers.Share which coding agent you usein the Privacy card of Settings. It applies to your account. It also stops the name of your agent.TROVE_NO_CLIENT_TELEMETRY=1on one computer. The CLI on that computer then sends no counts, no session key and no panel mark.
The server drops a count and the session key when one of these settings is off. After you change Usage data, the server can need up to 30 seconds to apply the change. Data that the server stored before you turned a setting off stays until you delete your account.
Direct CLI
You type these in Claude Code. Your agent cannot type them for you.
| Task | Command |
|---|---|
| Open or close the panel | /trove-panel |
| Change the settings | /trove-panel settings |
| Turn reading off | /trove-panel pause |
| Turn reading on (Auto) | /trove-panel auto |
| Turn the dashboard off | /trove-panel off |
| Turn the dashboard on again | /trove-panel on |
| Write, postpone, refuse or retry a book offer | /trove-panel book write, /trove-panel book later, /trove-panel book never, /trove-panel book retry |
| Write a subject from the first-run list, or hide the list | /trove-panel book write <n>, /trove-panel book close |
| List the commands | /trove-panel help |
Your agent runs these in a terminal:
| Task | Command |
|---|---|
| Check the dashboard | trove setup --status |
| Turn the dashboard off (from the next session) | trove mod settings set isOff true |
| Turn the dashboard on again | trove mod settings set isOff false |
| Remove Trove from Claude Code, the dashboard included | trove setup --disable claude-code |
| Put it back | trove setup --enable claude-code |
See trove setup and the Account API.