Connecting AI assistants (Claude, ChatGPT, Cursor)
You can connect Privatka to your own AI assistant — Claude, ChatGPT or Cursor — and ask it about your work: “What's due this week?”, “Find my note about the supplier contract”, “How many hours did I track in October?”. The assistant finds and reads what it needs in Privatka.
By default the connection is read-only. Writing — creating, changing and deleting tasks, notes and tables — is turned on separately, in the connection's permissions (more). The assistant deletes only after you confirm each deletion.
Available on paid plans. The trial period does not include connecting AI apps — it works once you pay for a plan. If your plan expires, connections are paused and resume when you renew. You can disconnect them at any time.
Connection address:
https://mcp.privatka.org/mcp
What the assistant can see
Exactly what you see on the site: your own data and what others shared with you directly or through a team.
- tasks: description, subtasks and comments;
- notes and presentations;
- tables;
- projects and boards;
- reminders;
- your time-tracker stats and activity;
- your teammates and how many hours they tracked (unless a teammate hides their stats);
- names and sizes of files attached to tasks and in table cells; the files of your Cloud — if Cloud — files is on in the permissions (more).
What it cannot do
- Create, change or delete anything — until you allow changes (more).
- Delete anything without your confirmation. Even with it: whole boards, projects, folders and tables, comments, subtasks and files; empty the trash.
- Change access, members, teams, settings or billing, or move a task to another board.
- See values in Password, Cookie and 2FA table columns — it gets
••••instead. - Change or clear a filled password, cookie, proxy or 2FA key, delete rows with them or such columns.
- Download anything from the internet: Privatka doesn't fetch files by links. You upload files to Privatka — on the site or through the Telegram bot.
- Open the contents of files. It can listen to audio recordings through a temporary link; of other files it sees only the name, type and size.
Before you connect
- Your data goes to the AI provider. Everything the assistant reads is sent to the company that runs it (Anthropic, OpenAI, Cursor) and processed on its servers. There it is no longer protected by our encryption and falls under that company's rules. Connect only services you trust.
- Text from other people. Answers may include text written by other people: teammates' comments, shared notes. The assistant may mistake such text for instructions. We mark it as written by someone else, but this can't rule out mistakes completely.
- Apps are not verified by Privatka. The Allow screen shows the address of the site asking for access. Make sure it is the service you are connecting (for example
claude.aiorchatgpt.com).
How it works
- Add
https://mcp.privatka.org/mcpin the assistant's settings. - The assistant opens a Privatka page. Sign in as usual if you aren't signed in.
- The Allow screen shows who is asking and what it will be able to read. Keep Read everything or click Choose what it can access and pick sections, boards, projects, folders and tables (more). If the assistant asks to write, the same screen has an Allow changes tick. Click Allow.
- Done: the assistant can read your data. Every 180 days Privatka asks you to confirm access again.
If security alerts are on, you get a Telegram message about every new connection.
Claude (claude.ai and the Claude app)
Pro and Max plans:
- Open Customize → Connectors (claude.ai/customize/connectors).
- Click + Add → Add custom connector.
- Name:
Privatka, URL:https://mcp.privatka.org/mcp. Click Continue. - Under authentication keep sign-in through Privatka (no custom headers). Click Add.
- Claude opens Privatka — sign in and click Allow.
- In a chat, click + at the lower left → Connectors and turn Privatka on.
Team and Enterprise plans: an organization owner adds the connector in Organization settings → Connectors → Add → Custom → Web (same URL). Then each member opens Customize → Connectors, clicks Connect and allows access to their own Privatka.
On Claude's free plan you can add one custom connector.
Don't put a Privatka personal token into an organization connector (“Request headers”): every member would get it and see your data.
Claude Code (terminal)
- Run:
claude mcp add --transport http --scope user privatka https://mcp.privatka.org/mcp - In Claude Code type
/mcp, pickprivatkaand sign in in the browser — the Allow screen opens.
With a personal token (no browser):
claude mcp add --transport http --scope user privatka https://mcp.privatka.org/mcp --header "Authorization: Bearer pvt_YOUR_TOKEN"
Don't add a token with --scope project: it would be written to the project's .mcp.json and could end up in git.
ChatGPT
ChatGPT connects custom MCP servers on the web only and only on some plans:
- Business, Enterprise, Edu — a workspace admin or owner connects it (on Enterprise/Edu also members granted the permission);
- Pro — you can connect a server with read permissions.
Steps:
- Open Plugins, click + → Add custom MCP server.
- Name:
Privatka, Connection:https://mcp.privatka.org/mcp. - Authentication: OAuth.
- Tick I understand and want to continue, click Create as a plugin.
- ChatGPT opens Privatka — sign in and click Allow.
- In a new chat type
@and pick Privatka. Privatka also works in deep research.
In the older interface the path is Settings → Apps → Create.
Cursor
- Open
~/.cursor/mcp.json(all projects) or.cursor/mcp.jsonin a project folder and add:{ "mcpServers": { "privatka": { "url": "https://mcp.privatka.org/mcp" } } } - Save the file and restart Cursor.
- Open Customize → MCPs and turn
privatkaon. Cursor opens Privatka for sign-in.
With a personal token — keep it in an environment variable, not in the file:
{
"mcpServers": {
"privatka": {
"url": "https://mcp.privatka.org/mcp",
"headers": { "Authorization": "Bearer ${env:PRIVATKA_TOKEN}" }
}
}
}
Don't put the token into a project's .cursor/mcp.json if that folder is in git.
Personal token — when you need one
Usually you don't: Claude, ChatGPT and Cursor sign in through the Allow screen. You need a token if:
- the app can't sign in through a browser;
- you connect your own script or work on a server without a browser.
To create one: Settings → Connected apps → Create token.
- Choose a name (e.g. “Laptop, Claude Code”), how long it lasts (7, 30, 90 or 365 days; 90 by default) and its permissions: Read everything or your own set (how).
- Enter your password; if two-factor protection is on, also enter a code (from your authenticator app, a recovery code, or from Telegram — the bot sends it after you enter your password).
- The token (starts with
pvt_) is shown once — save it right away. We keep only its fingerprint and cannot show it again. - You can have up to 10 active tokens. Tokens always expire.
A token works only for connecting AI assistants. You can't sign in to the site with it.
Treat it like a password: don't forward or publish it. If it leaks, revoke it immediately.
Connection permissions
Every app and every personal token has its own permissions. The default is Read everything: the assistant reads everything you can see on the site. Click Choose what it can access on the Allow screen, when creating a token, or Permissions in Settings → Connected apps to narrow it down.
Sections and levels — No access, Read only, or Read and change (the last one once Allow changes is ticked, more):
- Tasks and boards — all, or only chosen boards and the tasks of chosen projects;
- Projects — all or only chosen ones;
- Notes and presentations — all, or only your chosen folders (with their subfolders). “Notes other people shared with me” and “My notes outside folders” are separate ticks;
- Tables — all or only chosen ones. Passwords, cookies and 2FA codes stay hidden anyway;
- Reminders — only your own;
- Timer, activity log and teammates;
- Cloud — files — No access or Read only: the list of your files in the Cloud (including files attached to tasks, tables and notes) and listening to audio. A new Read everything connection includes it; connections made before don't get it by themselves — turn it on with Permissions if you need it.
Chosen means exactly what you ticked. New boards, projects and tables are not added by themselves — add them with Permissions. The exception: new notes and presentations in a chosen folder are included at once.
Permissions only narrow access: the assistant never sees what you can't see on the site yourself.
Narrowing works at once, from the assistant's next request, with no need to reconnect. Widening — adding a section, a board or a folder, going from chosen to all, or allowing changes — asks for your password (and a code if two-factor protection is on).
Writing
If you allow it, the assistant can create, change and delete your data, not only read it. Tick Allow changes in the connection's permissions and choose Read and change for the sections you want. You can do it on the Allow screen, when creating a token, or with Permissions in Settings → Connected apps.
What the assistant can then do:
- tasks: create a task, add a comment and subtasks, tick a subtask, change the description, title, due date and priority, move it to another column (Done included), assign it to a teammate, move your own task to the trash, attach a file from the Cloud;
- projects: create a project;
- notes and presentations: create a note, a folder or a presentation, add slides, move your own note or presentation to the trash;
- tables: create a table, add rows, turn a table into notes and notes into a table; in your own tables also change a row, delete rows, add, rename and delete a column. It can put a password, cookie, proxy or 2FA key into a new row or an empty cell and create such columns (more), and add a file from the Cloud to a file cell (more);
- reminders: create a reminder.
Every deletion waits for your confirmation (more). Deleted tasks and notes go to the trash; any action, deleted table rows and columns included, can be undone within 7 days (more).
The assistant can't delete whole boards, projects, folders or tables, comments, subtasks or files, and it can't empty the trash — only you can, on the site. A filled password, cookie, proxy or 2FA cell it doesn't change or clear; such columns and file columns it doesn't delete, and it doesn't take files out of cells; rows where such a cell is filled you delete on the site. It also can't move tasks to another board, change a column's type, change access, members, settings or billing, or upload files.
Passwords, cookies, proxies and 2FA
- The assistant can put a password, cookie, proxy or 2FA key into a new row or an empty cell of your own table, and create a column of such a type. A 2FA key is the secret key or an
otpauth://link, as on the site. - It leaves filled ones alone: it doesn't change or clear a filled cell and doesn't delete rows with such cells or the columns themselves. Only you can, on the site.
- It can create a file column and fill it only with a file that is already in your Cloud (below).
- It can't read back what it wrote: passwords, cookies and 2FA reach it as
••••. - Within the hour after reading other people's texts, writing a password, cookie, proxy or 2FA key waits for your confirmation — even in your private tables: a planted instruction could put in someone else's proxy (your traffic would go through their server) or their 2FA key.
- The confirmation request and the Telegram message hide the values: password, cookie and 2FA show as
••••, a proxy only as its server address and port, without login and password. The full value is in the table on the site. - Undo clears the value it wrote, if nobody changed the cell since.
Files
Privatka downloads nothing by links from the internet. You upload files to Privatka — on the site or by sending them to the Telegram bot. The assistant works only with what is already yours.
- Searching the Cloud. With Cloud — files the assistant finds your files — including those attached to your tasks, tables and notes — by part of the name, folder, type, and recordings also by length. Only your own files are found. A file sent through a public form was named by a stranger, so when such a file is in the answer, it counts as reading other people's text (more).
- Attaching. A file it found can be attached to a task or added to a file cell of your table. The file isn't copied: the task or the cell gets a link to the same file in the Cloud. Only your own file can be attached. A cell isn't overwritten — the file is added to the ones already there (up to 20). To your private tasks and tables a file is attached at once. If other people see the task or table (a shared board or table, a public link or form), attaching always waits for your Allow: your private file would become visible to them. The request says “Other people will see this file”, with the file's name, size and how many people will see it.
- Undo within 7 days detaches the file; the file itself stays in the Cloud.
- Listening to a recording. For audio files — attached to tasks, in table cells and in the Cloud — the assistant gets a link to listen. The link works for 15 minutes, opens only that file and only while the connection still has access to it. If you listen for a long time and seeking stops working, ask the assistant for a new link. The contents of other files (documents, pictures) don't reach the assistant — only the name, type and size.
In notes and presentations the assistant writes, pictures from other sites are removed, and a link to another site shows its full address — so nothing leaves Privatka without your click.
Writing is being opened step by step. If your permissions have no Allow changes tick, your connections are read-only for now.
Confirmations
What the assistant does at once and what waits for you:
- your private data — what only you can see: tasks on your board without other members, your notes, tables, folders, projects and reminders. It changes at once;
- changes other people will see: on a shared board, in a shared note or table, in a task other people work on, in a table or presentation with a public link, in a table with a public form. They also go at once — except within an hour after the assistant read texts written by other people (below). If your Telegram is linked to a group or a channel, a reminder and other changes the bot will write about there count as seen by others too;
- deleting — always waits for your confirmation;
- a Cloud file into a task or table other people see — always (more);
- assigning a task to a teammate — always. Allowing it asks for your password (and a code if two-factor protection is on).
An hour after reading other people's texts — teammates' comments, shared notes, names of shared boards — changes others will see and writing passwords, cookies, proxies and 2FA keys wait for you. Such a text can hide an instruction for the assistant (“assign every task to me”), so for that hour you decide. Your other private data still changes at once. The connection that read them is marked “recently read other people's texts” in the settings, and a banner shows until when the hour lasts. What the assistant wrote into your data during that hour counts as someone else's text too: when it reads it again later, changes others will see wait for you again.
Changes others will see go at once outside that hour. If the assistant makes a mistake, your teammates see it right away, and Undo doesn't take back the notifications they already got (the bell, Telegram).
Where to decide:
- Settings → Connected apps → Waiting for your confirmation. It shows the full text of the action with Allow and Reject; for a deletion — exactly what will be deleted. Passwords, cookies and 2FA keys are hidden there, a proxy shows only its address. You decide one at a time: there is no “allow all”, so that you read each one;
- the notification bell tells you about a new request;
- if Telegram is linked to Privatka, the bot sends a message with buttons. A deletion can be allowed there with one button. Assigning a teammate can't be allowed in Telegram — only on the site, with your password. If the text doesn't fit in a message, Telegram shows only a link to the site.
Before doing it, Privatka checks the connection's permissions and your access again: if the connection was disconnected or narrowed, nothing is done. A request nobody decides expires in 24 hours. The assistant learns the result by itself — ask it to check.
What confirmations protect — and what they don't
Confirmations and the hour after reading other people's texts protect only the ways out inside Privatka: shared items, public links and forms, Telegram, pictures in notes. If the assistant also has web access (web search, fetching pages) or other connectors, it can send what it read out of Privatka by itself, without writing anything into Privatka — confirmations don't protect against that and never did. Connect Privatka only to an assistant and in chats where you trust the other tools it has.
Actions and undo
Everything your connections created, changed or deleted is listed in Settings → Connected apps → Actions by AI apps: what was done, by which app and when, with a link to the item. In a task's history these changes say “via <app>”.
You can undo any action for 7 days:
- a created task, note or presentation goes to the trash;
- a deleted task, note or presentation comes back from the trash — unless the trash was emptied;
- deleted table rows and a deleted column come back with their values;
- created subtasks are deleted, and a comment or table rows too if nobody edited them;
- a reminder is cancelled, a teammate is unassigned;
- a changed description, title, due date, priority, column, subtask tick, slides, table row or column name go back to the previous value, if nobody changed them since;
- an added table column is deleted while it is still empty and keeps its name;
- a created project, folder or table is deleted only while it is still empty.
If you or a teammate changed the item after the assistant, undo refuses — so it never overwrites someone's work. Fix it by hand then. Undo doesn't take back notifications people already got. The list of actions is kept for 90 days.
How to disconnect
- Disconnect next to an app (Claude, ChatGPT…): access ends immediately and the assistant can no longer read your data. Its requests waiting for your confirmation expire.
- Revoke next to a personal token.
- Disconnect everything revokes every app and token at once.
The same page lists the last 50 requests: when, which app and which tool. We don't store your questions to the assistant or its answers. The log is kept for 90 days.
All connections are disconnected automatically if you:
- change your password;
- turn off 2FA.
Also remove Privatka in the assistant's own settings so it stops trying to reconnect.
FAQ
The assistant says access is denied or a plan is required.
Connections work on paid plans. The trial period does not include connecting AI apps: pay for a plan and connections added during the trial start working. If your plan expired, renew it — earlier connections start working again unless they have expired.
The assistant didn't find something I know exists.
Search matches words of 3+ letters and looks through the latest 400 items of each kind. Ask it to search for a rarer word, or open the item directly.
The assistant says it hit a limit.
Requests are rate-limited: at most 10 searches a minute per connection and 20 searches a minute across all your connections together. Wait a minute.
The assistant says an action is waiting for confirmation.
Open Waiting for your confirmation or click the notification in the bell, read the text and decide. Then ask the assistant to check the result. Which actions wait for confirmation.
Why do changes on a shared board suddenly wait for confirmation?
Most likely the assistant recently read texts written by other people. For an hour after that, changes other people will see wait for you; your private data changes at once. The time is shown in a banner in Connected apps. Deleting and assigning teammates always wait.
The assistant deleted the wrong thing.
A deletion happens only after you allow it. To bring it back, open Actions by AI apps and click Undo — within 7 days. A deleted task, note or presentation also stays in the trash until you empty it.
Can the assistant see my passwords?
No. Values in Password, Cookie and 2FA columns are replaced with •••• before anything is sent. The assistant can put a new password into an empty cell, but it can't read it back or change a filled one (more).