# Is it safe to give ChatGPT access to your notes? You have seen what a connected knowledge base can do: ChatGPT that actually knows your projects, remembers decisions, and saves what you tell it. The reason you haven't set it up is a different question: what exactly am I giving it access to, and what happens when it gets something wrong? Fair question. It deserves a specific answer, not a shrug.

Quick answer

Connecting ChatGPT to your notes is safe when three things are true: the connection is revocable in one click, every edit is versioned and attributed so you can see and undo anything the AI changed, and the knowledge base itself does not train on your data and exports to plain Markdown. The real risk is not the AI. It is using a tool where you cannot see what changed.

## What a connected assistant can actually do A connection over MCP is not "ChatGPT gets your password." It is a revocable session, tied to your account, that exposes a specific set of tools. In Hjarni's case, a connected assistant can: - Read notes, folders, tags, and AI instructions in the spaces your account can access. - Create and update notes, tags, links, and folders. - Work in team spaces you belong to. And it cannot: - Touch anything before you connect it. - Bypass ownership or sharing rules. - Reach spaces your own account cannot reach. - Keep memory anywhere outside the notes and instructions you can see. One honest limitation: an OAuth MCP connection reaches everything your account can reach, and it cannot currently be limited to a single folder. If you want narrower access for a script or an integration, [API tokens can be scoped](/docs/privacy-and-permissions) to one folder, your personal notes, or one team. Either way, revoking from Settings cuts access immediately. ChatGPT adds its own layer on top: at the time of writing it asks you to approve writes and shows you the payload before running them. That is ChatGPT's behavior, not ours, so treat it as a bonus rather than the guarantee. The guarantee is the part we control: everything below. ## The three worries, taken seriously **"It will make a bad edit."** It might. So will you. What matters is whether you can see it and undo it. In Hjarni, [every note keeps a full, attributed history](/changelog/note-history-and-provenance): AI edits carry a badge, every revision shows exactly what changed, and one click reverts to any earlier version. Deletes go to Trash and stay recoverable for 30 days. A bad edit costs you a click, not your notes. A bad run of them costs you a few clicks: history is per note, so undoing a wide sweep means walking the notes it touched. **"My data leaves my control."** Two separate things hide in this worry. Hjarni does not train models on your data, full stop. Separately, whatever an assistant reads from your notes becomes part of that conversation, like anything you paste into it, governed by the data settings you already chose for ChatGPT or Claude. Connecting a knowledge base does not change those settings. It changes what the assistant can look up while you work. **"I'll get locked in."** The exit is the test. Your entire knowledge base [exports as a ZIP of Markdown files](/docs/export-and-ownership) with the folder structure preserved. If you leave, you leave with everything, in a format every tool on earth can read. ## What the data says about how AI actually behaves We measured it. In the last 30 days, assistants made [108,613 tool calls against Hjarni knowledge bases](/blog/what-ai-does-with-write-access). Over half were reads. Assistants fetched the user's custom instructions 9,480 times before acting, checked note history 2,320 times, and updated existing notes 3.7 times for every new note they created. Just 0.9% of calls returned an error. Connected to a knowledge base with structure and rules, the AI behaves like a careful co-author, not a bull in your archive. ## How to connect safely in five minutes 1. [Create a free account](/registration/new) and save a handful of notes. Nothing sensitive on day one if you'd rather warm up to it. 2. Connect ChatGPT from the Apps Directory, or follow the [setup guide](/docs/connect-chatgpt-mcp). Claude works the same way over MCP. 3. Ask it to search your notes and save something. Then open the note's history and look at the revision it just wrote: attributed, diffed, revertible. 4. When you trust the loop, move your real knowledge in. That last step is the whole point. The safeguards are not there to make AI access tolerable. They are there so you can stop re-explaining yourself and let your AI actually use what you know.