What leaves your machine

Last reviewed: 31 Oct 2026

Before you install anything, you should know what it sends and to whom. This page is the whole list. Everything here goes to your own Naetive account, over an authenticated connection, and nothing else.

A chat agent (ChatGPT, Claude on the web, Claude Desktop)

Nothing is installed. The agent calls Naetive when it uses a tool, and what it writes (a job, an update, an entry) is stored. Naetive never sees the rest of your chat; the agent's provider handles that as it always did. Not connecting the agent is the opt-out.

Claude Code, installed from Connect

The installer puts five hooks into Claude Code. Each one sends exactly this:

Activity. After Claude Code runs a tool, the hook sends the tool's name, the file path, the command or the search pattern, whether it worked, and for a command up to the first two kilobytes of its output, tagged with the window it came from. This is what shows a window as live now on Work with the files it touched. It reads only what Claude Code already handed it; it does not scan your disk. When the call changed the understanding, it also refreshes the .naetive folder in a repository that has one; that part only fetches.
Session start. When a window opens, the hook prints which project this folder is linked to (on this computer, for the agent), checks for an update to the naetive command at most every six hours, and catches up the .naetive folder in a repository that has one. It fetches; it sends nothing about your work.
Chat link. When a chat started from the Chats page opens, the hook tells the console on your machine which chat is in which terminal (the chat's id), and fetches a short project brief for the agent.
Prompt grounding. When you send a prompt of 12 characters or more, in a folder linked to a project or a chat started from Chats, the hook sends the first 300 characters to Naetive through the local console, to fetch the recorded work that matches it (jobs, decisions, specs) and put it in front of the agent. Shorter prompts are skipped. If you do not want your prompts to leave the machine, remove this hook (below) and the agent still works; it just orients by itself.
Key guard. Runs before a read, a search or a command and blocks any that would print your key. It sends nothing.

On Windows the hooks are not installed yet. Claude Code connects and works, and the console and tunnel run and report as described below; but a window's file edits are not reported and no prompt text leaves the machine.

The tunnel

The tunnel is a program on your computer that runs only while this computer shares a project (it starts by itself in the background; on Windows you start it yourself). While it runs, agents that live elsewhere (ChatGPT, Claude on the web) can read and change the projects you shared on that computer, and only those, through Naetive, acting as you. It refuses a request for any other project. Stop it and the access ends. It sends file contents and command output only when an agent asks for them, and only from those folders. When it starts it tells Naetive which computer it is, which projects it shares and their folder paths, so Connect can show them. If the console has crashed several times in a row and given up, the tunnel tells Naetive how many times and when, nothing more.

The Chats page and the console

The console is a small program on your machine (port 4000) that runs the real agent process when you start one from the Chats page, and streams it to your browser. It binds to your own machine only and answers only to the Naetive that installed it. While it runs it reports in every minute with this computer's name and its own version, so Connect can show it running. What the agent does in that session is reported the same way as any Claude Code window.

Chats history

So History works across your computers, Naetive stores one line per chat: the computer's name, the chat's folder (shown as ~/... when it is under your home folder), when it ran, and the chat's name and first line, with anything that looks like a secret removed. It does not store the chat itself. It is written from what your own tunnel reports, only for your account. You can hide a chat from History or delete it for good.

The `naetive` command

naetive export fetches the understanding and writes it into the repo. It sends nothing except your key to authenticate. naetive share and naetive unshare also send a fingerprint of this computer's shared list (not the list of files), so Naetive can tell if it was changed some other way. naetive commands on sends this computer's name, to show on the page where you confirm.

Health reports

Check and fix (naetive doctor) looks over the Naetive pieces on your computer every 10 minutes and repairs what it can. When it finds something it can't fix, the computer sends Naetive a short health report so we can help before you have to ask. A report holds exactly this, and the server refuses anything else:

which check (one of: the key, the console, the tunnel, a second copy of the tunnel, start at login, Claude Code's connector and hooks) and whether it failed or was fixed;
the version of the console, the tunnel and the naetive command;
the kind of computer (Linux, macOS, Windows or WSL), its main version number and chip type, and the Node version;
the time.

Never your files, folder paths, file contents, error messages, computer or user name, project or chat names, or your key. It is sent with this computer's own key, at most once per check every six hours, and only when something changed. Only Naetive's admins see reports, and they are deleted after 30 days. Agents never see them.

It is on by default. Turn it off with Send health reports to Naetive on Connect, under Settings: the computer stops sending and Naetive stores nothing. To stop it on one computer only, run naetive health-reports off there. Send to Naetive on a problem Check and fix couldn't fix is separate: it sends only when you press it, and only there can you add a note in your own words.

Your key

When you approve this computer during install, it gets its own key, and the installer writes it to a file under ~/.claude, readable only by you. Nothing puts it on a command line, and the key guard stops an agent from printing it. Chat agents and agents set up by hand get their own key too, shown once when you make it on Connect; it travels inside the connector URL or config you paste, so treat that URL as secret. Your account also has its own key, shown in full on Connect only once, right after you make it. Naetive stores only a fingerprint of each key. If the account key may have leaked, Regenerate on Connect: the old key stops working at once, including a tunnel or terminal already running on it, and the key card lists what to reconnect, ticking each step off as it happens. Computers and agents with their own key are not affected. If an agent's own key may have leaked, make a new one for that agent; the old one stops working. If a computer's key is revoked, its tunnel stops and says so.

Switching things off

One folder: mkdir -p .claude && touch .claude/naetive-disabled in that folder, in a terminal on macOS, Linux or WSL (there are no hooks on Windows). The activity and session start hooks stay silent there; not even file paths are sent. Prompt grounding does not look at this file: to stop it, remove that hook.
One hook: remove its line from Claude Code's settings (~/.claude/settings.json). Claude Code stays connected.
Health reports: Send health reports to Naetive on Connect (Settings), or naetive health-reports off on one computer.
Everything: the uninstaller removes all of it, including the connection. See Removing Naetive from a computer.

What Naetive does with it

It shows it to you, and to the agents you connect, on your project. Naetive makes no model calls of its own over it; the understanding is assembled from what agents recorded, not generated. It is your account's data and nobody else's.