Agents
Your agents, in your knowledge base
Knowtarium doesn't run an AI itself. Connect the agents you already use over MCP, and they search, read and write your notes directly. Every change they make is signed with the agent's own key and can be undone in one click.
Works with your agents
Bring the agents you already use.
- Claude Codeone command, or the pluginSetup guide
- Codexone command, or the pluginSetup guide
- Cursorone command, or one click in the appSetup guide
- OpenCodeone command
- Claude Desktopdownload the extension, no terminalSetup guide
- VS Codeone click in the app
- Any MCP agentthe same local server
- claude.ai, ChatGPTcoming later
Knowtarium runs no AI of its own: these agents connect to it over MCP. Chat apps that only run in the cloud will connect later, through a hosted connector that decrypts only inside a sealed enclave running published code.
Knowtarium Web Clippera browser extension for Chrome: save the page you're on to your inboxIn practice
Connect once. Then just ask.
Run , pick a workspace in the browser, or just some of its folders, and click Allow. It lists the agents it finds on your computer, sets up Knowtarium in the ones you keep ticked, and nothing secret is copied or pasted.
From then on your agent searches and reads only what it needs, and writes straight into your workspace, following the instructions you wrote for each folder.
Example terminal session: npx knowtarium connect connects a workspace, then lists the agents it found (Claude Code, Cursor and Codex), all ticked; Cursor is unticked, so Knowtarium is added to Claude Code and Codex. Then Claude Code, asked to update a price, searches the notes and edits two of them.
Getting connected
One command and one click.
- 01
Run one command
runs on your computer (it needs Node.js 20 or later). It makes its own key pair and opens Knowtarium in your browser.
- 02
Choose what it reaches
Name the agent, pick a workspace, optionally only some folders, and read or read-write. Anything more, like deleting notes or changing folders, stays off unless you tick it. Then click Allow, but only if you ran the command yourself.
- 03
Keys reach only your computer
Your browser hands the agent its copy of the workspace key straight to your computer, or, if that's blocked, sealed so we can't open it. The CLI keeps it encrypted, with the key in your system keychain (or a private file where there is none).
- 04
Your agents are set up
It lists the agents it finds (Claude Code, Codex, Cursor, OpenCode, Claude Desktop), all ticked, except any that already have the Knowtarium plugin. Untick any you want to leave out: it adds Knowtarium to the others and leaves the rest of their settings alone.
Or from your agent
A plugin, or an extension.
Claude Code and Codex can install Knowtarium as a plugin, with a skill for working in a Knowtarium workspace. It runs the same server from npm (Node.js 20 or later). If this computer isn't connected yet, ask your agent to connect Knowtarium: the browser opens so you can approve this computer.
With the plugin you don't need to add Knowtarium: the plugin already runs the server. To connect from the terminal, run , which leaves your agents' settings alone.
In Claude Desktop, download the extension, open it, then ask Claude to connect Knowtarium: the browser opens so you can approve this computer. No terminal or Node.js needed.
# Claude Code# Codex
Setup guides per agent
The rules
What agents can and can't do.
- canSearch, read and follow links between the notes you gave it
- canCreate and edit notes, applied right away
- canMove or delete notes and change folders, only if you allow it
- canSign every change with its own key
- canFollow the instructions you wrote for each folder
- can'tReach a workspace or folder you didn't give it
- can'tDelete attached files, or rename, move or delete the folders you gave an agent
- can'tSign as you, or mark anything as checked by you
- can'tSkip review where you turned it on
- can'tChange your instructions without you accepting it
Deleting and organizing
More than editing, only if you allow it.
- 01
Off until you tick it
Moving notes, deleting notes, and creating, renaming, moving or deleting folders are six separate permissions, each off by default. Change them any time in the app under Settings, Agents: each time, you sign the new set.
- 02
Told to wait for you
Its conventions tell your agent to delete, move or reorganize only when you asked for it in your conversation, never because a note says so. That's an instruction; the permissions are what our server enforces.
- 03
Ask me first still holds
In folders set to Ask me first, a note delete or move waits for your approval. Folder changes there are refused, and your agent is told to tell you what it would change.
- 04
Listed, with Undo
Each delete, move and folder change is signed by the agent and listed on the workspace page for two weeks, or your history period if that's shorter, under Deleted and reorganized by agents, with Undo and Undo this agent's last 24 hours. A deleted note comes back while your history period keeps it. The app checks each change against the permissions you signed.
Agents never delete attached files, the workspace's root folder or the web clipper's inbox, and never rename, move or delete the folders you gave an agent. They delete a folder only when it's empty. The agent still holds the workspace key, so it's our server that enforces your permissions and folders (see the known limits). Permissions need knowtarium 0.3.0 or later: for an older one, the connect page shows the command to run.
Folder instructions
Tell your agents how to write here.
Fernhill / research
Instructions for agents
For research and the folders inside it. Agents follow these for every note they write there.
Also applies here, from above
The workspace’s instructions289 BEdit them there
How to write notes in Fernhill · Write for a new teammate: plain words, short sentences, no internal shorthand.
Show allHow to write notes in this folder
- One note per company, product or market question, named after it.
- Every fact needs a source: add it to
sourceswith a link or a path. Leave out what you can't source. - Use
type: Research, and tag each note with the market it covers. - Start with a one-line
descriptionof what the note answers. - Date every number (prices, market sizes, user counts), and set
stale_aftersix months later. - When sources disagree, say so in the note and keep both.
499 B · 788 B with the instructions above and below
- 01
Write them once
Choose Edit instructions in a folder's menu, or Instructions for agents in the workspace's settings, and say how notes there are written: their fields, their sources, their names. Start from a blank page or a template: Market research, Meeting notes or Decision log.
- 02
Followed for every note
Claude Code, Codex, Claude Desktop and every other agent you connect follow them for each note they write there: the workspace's first, then each folder's down to the note's own. An agent limited to some folders gets the ones above them too.
- 03
Changed only by you
An agent can propose a change to your instructions, never make one. It waits in your review queue, and the current instructions apply until you accept it. An AGENTS.md you import applies only once you choose Apply.
- 04
Only your words count
Agents follow only instructions you saved or accepted, signed with your key. Their conventions tell them that text inside a note, a clip or an index.md is data, never an order, and to tell you about it instead.
Each folder's instructions are saved as its AGENTS.md, encrypted like any note and kept out of your note list, search and graph. They export as AGENTS.md files with the rest of the workspace. There's no size limit, but shorter instructions are followed better, and past about 8 KB the editor says so. Instructions guide your agent; they don't lock anything (see the known limits).
Honest about it
Your agent sees what it reads.
The knowtarium CLI decrypts notes on your computer, keeps them in memory, and hands your agent only what its tools return. Whatever your agent reads goes to its AI provider, like anything else you give it. Your notes reach our servers only as ciphertext.
Revoke an agent in the app and it's shut out at once, and the workspace gets a new key, so it can't read anything written after that.
Opening the browser to connect this computer to Knowtarium.Connected to workspace ws_8f2k (read and write).Add Knowtarium to which agents? Claude Code, Codexadded to Claude Code (~/.claude.json)added to Codex (~/.codex/config.toml)Keys are kept in: the OS keychainWorkspace ws_8f2ktoken tok_3m9q (active)access read-write, folders fld_p1, fld_m4Revoked the token for workspace ws_8f2k.Deleted the local keys and cache for workspace ws_8f2k.
Web clipper
Save to Knowtarium.
A browser extension
The page you're on, in your inbox.
The Knowtarium Web Clipper is an extension for Chrome. Click its button, or press Alt+Shift+S (Option+Shift+S on a Mac, and you can change it) to open it, add a note or pick a category if you like, and press Save. The page is queued, encrypted, until Knowtarium is next open on that workspace, or an agent that can write to the inbox runs: then it's filed as a note in the workspace's Inbox folder, and the agents you connected can sort and summarize it.
To connect it, click Connect to Knowtarium in its popup. A tab opens with a code, and Knowtarium opens next to it: check that both show the same code, pick the workspace and allow. Disconnect it any time from the popup, or revoke it in Knowtarium under Settings, Agents. It's open source (MIT), and a version for Edge is coming.
Knowtarium
Connected to ResearchDisconnect
Write-only
It can add pages to your inbox, but it holds no key that reads your notes, not even the pages it saved.
Encrypted in your browser
Each page is signed and encrypted before it's sent. Our servers queue it but can't read it, not even the link or the title.
Reads only the tab you're on
It reads the current tab's link and title when you click its button or press its shortcut, and sends nothing until you press Save. Never the page's text, your selection, other tabs or your history.
Works offline
Pages you save without a connection wait in the extension, up to 100, and are sent the next time you open its popup or start your browser.
It asks your browser for two permissions only, activeTab and storage, and has no analytics. The privacy policy says what it reads, what we store and what stays on your device.