Connect your AI (MCP)

Your own AI app, such as Claude, ChatGPT, Claude Code or Cursor, can fill and maintain your knowledge base for you: move articles over from Zendesk or Notion, upload a folder of documents, remove what disappeared from your site, and check that the bot finds the answers. You connect it on Admin > AI > Connect your AI by pasting one address into the app and clicking Allow in your panel; no keys are involved.

How it works

Every platform has its own knowledge server, which speaks MCP, the protocol AI apps use to connect to outside tools. Your AI app does the reading: it opens the source, whatever it is, and turns it into articles. The knowledge server makes sure the writing is safe. Files and web pages are processed by your panel exactly as when you upload them yourself (see Importing files, links and sitemaps).

A Confluence wiki is better connected directly on the knowledge base page (see Confluence connector): it then updates by itself, and its articles cannot be changed through the knowledge server.

What is guaranteed

• Plan first. Every write can be run as a dry run, and the AI is told to show you what it will add, update and delete before it writes anything.
• Deletions and rollbacks always go through a plan. The AI first shows what will change, and the plan must be confirmed within 10 minutes. Deleting more than 25 articles and more than a fifth of the knowledge base at once needs one more explicit confirmation.
• Your edits are never lost silently. If an article was edited in the panel and its source has changed since, the AI gets a conflict instead of overwriting it and has to ask you.
• No duplicates. Every article the AI writes carries a key from its source, so importing the same source again updates the same articles.
• Imports can be undone. Each import and each deletion made through the knowledge server is recorded as a batch that can be rolled back for 30 days. An article changed after the batch is reported as a conflict instead of being rolled back silently.

Connect Claude, ChatGPT, Claude Code or Cursor

On Admin > AI > Connect your AI, copy the Knowledge server address. The page has a tab with steps for each app:

• Claude (web, desktop or mobile app): Settings, Connectors, Add custom connector. Paste the address, add the connector and click Connect. Your panel opens: sign in if asked and click Allow. On Claude Team and Enterprise plans an organization owner adds the connector.
• ChatGPT: in the settings open Apps & Connectors and turn on developer mode in the advanced settings. Create a connector with the address, choose OAuth authentication and sign in to your panel when ChatGPT asks.
• Claude Code: run the command shown on the page (claude mcp add --transport http kb followed by the address), then type /mcp in Claude Code, pick kb and choose Authenticate.
• Cursor: add the address to Cursor's mcp.json, as shown on the page. When Cursor connects, it opens your panel to sign in.

Who can allow a connection

The app opens a page in your panel titled Access to the knowledge base. It names the app, where the browser returns after you allow, what the app gets (reading, adding, changing and deleting knowledge base articles on your behalf, and nothing else in the admin) and which account you are signed in with. If the page says the app's name is not verified, check the return address before you go on. Click Allow to connect or Deny to refuse. The request is valid for 10 minutes; after that, start connecting again in the app.

Only a person with the Manage AI features permission can allow a connection: for anyone else the page says so and Allow stays unavailable. The app then acts on behalf of the person who allowed it and reaches nothing but the knowledge base.

Connected apps

The Connected apps list on the same page shows every app someone allowed, who allowed it, when it was connected and when it was last used. Disconnect cuts the app off at once; next time it will ask to sign in again. A new password for the person who allowed the app does the same for all of that person's apps.

Access also ends immediately when the person who allowed it is disabled or loses the Manage AI features permission.

What your AI can do

Through the knowledge server your AI app can:

• show the state of the knowledge base: how many articles are searchable, which failed to index, recent batches;
• list articles and read any of them;
• create or update articles from text it prepared, for example from Zendesk, Notion or a document it read, up to 50 per call;
• upload files from your computer (PDF, Markdown, text, HTML, up to 25 MB each) through a one-time upload link;
• import web pages or the pages of a sitemap from public addresses, showing the list of pages first;
• delete articles, or mirror a source by deleting what disappeared from it;
• reindex articles that failed to index;
• export the knowledge base as a zip file of Markdown articles;
• ask your assistant a question the way a customer would and see its reply and the articles it found, the same check as the Test tab of the assistant;
• roll back a batch.

Uploading local files needs an app that can run commands on your computer, such as Claude Code or Cursor. Claude on the web or desktop and ChatGPT cannot do that: they send text, and files you upload in the admin yourself.

What to ask

The page suggests requests you can copy, for example:

• "Move all articles from our Zendesk help center into the knowledge base. Show me the plan first."
• "Upload the documents from the “Guides” folder to the knowledge base."
• "Compare the knowledge base with our site's sitemap and delete articles that are no longer there."
• "Check that the bot finds the answer to “how do I return an item”."
• "Roll back the last import."

To check the whole knowledge base at once, use the ready request in the block 4. Check the knowledge base on the same page, or the Check the knowledge base prompt in Claude, see Check your knowledge base with your AI.

Limits

• One write at a time per knowledge server.
• A request carries up to 2 MB of text; a file up to 25 MB.
• One sitemap import takes up to 300 pages; other pages can be passed as lists of addresses, up to 50 per call.
• Long jobs, such as a big sitemap, run in steps of about 45 seconds, and the AI continues until they are done.
• An answer to the AI is up to 40,000 characters, so long articles are read in parts.
• Questions to the assistant: up to 10 a minute.

An API key for scripts

Scripts, CI jobs and tools that cannot sign in through a browser use an API key instead. In the section For scripts: an API key, enter a Connection name and click Create key. The panel creates a role "KB import (MCP)" with the Manage AI features permission, a separate agent with that role and no teams (so no conversations are assigned to it) and a key, then shows a ready Claude Code command with the key. Copy it right away: the secret is not shown again.

Creating a key needs the rights to manage roles and agents. Without them, ask an administrator, who can also do the same by hand: Roles, a new role with Manage AI features; Agents, a new agent with that role and no teams; then Generate API key in the agent's card. The agent made by Create key does not count toward the agents in your plan; an agent made by hand counts like any other enabled agent.

Unlike a browser connection, a key also opens the AI settings, providers and assistants included. Disable the agent or revoke its key, and access ends at once. In Claude Code, keep a server added with a key under a different name from the one that signs in through the browser (the command on the page names it kb-key): with a key header, Claude Code switches the browser sign-in off.