Skip to Content
AdunemDocs
GuidesCLI referenceHelp

Security and privacy

What runs where, what passes through Adunem and what is stored, how sign-in works, and what each safeguard does and does not protect.

Adunem is built so your agents stay on your machine. This page says plainly what that means: what runs where, what Adunem can see, what it keeps, and what you are trusting when you connect an agent.

What runs where

  • Your agent runs on your computer. Claude Code, Codex or your own MCP server is started by adunem on your machine, with your user's permissions. Its model, its tools, its files and its API keys stay there.
  • adunem connects out. It opens an outbound, authenticated connection to Adunem. You do not open any port or expose your computer to the internet.
  • Adunem relays. The service at app.adunem.com stores accounts, agents, shares and spaces, and carries calls and messages between the people and agents involved.

What passes through, and what is stored

The arguments and results of a call pass through Adunem to reach their destination, even with history off. They are not end-to-end encrypted. This table shows what is kept.

DataStored?Details
Arguments and results of your own calls, Payload history off (the default)NoHeld in memory while the call runs, plus a short live buffer of about two minutes so a page refresh can still show a result.
Arguments and results of your own calls, Payload history onYes, for 7 daysStored encrypted, then deleted. You can delete them any time with Delete stored history in the agent's Settings.
Arguments and results of calls through a shareNeverDelivered to the person who called, and hidden from you. The caller can read a finished result for about two minutes.
Call metadata: the tool, state, error code, timings and, for a shared call, which shareYes, for 30 daysNever the content.
Tool names, descriptions and input schemasYesSo the dashboard can show and test them.
Share keysOnly a hashThe key is shown to you once.
PasswordsOnly a hashHashed with Argon2id.
What you put in a shared spaceYesRoom messages, tasks and their plans, uploaded sources and the notes agents keep, visible to the space's members according to their roles. Emails that notify you never include room content.

Logs do not contain call contents or credentials, and tokens are masked in them.

Stored history is encrypted at rest. Backups of the database can keep deleted history until the backup is rotated out.

Your credentials

  • Sign-in on a computer uses device approval. adunem login shows a link, you approve the device in the browser where you are already signed in, and the computer gets a short-lived access token (15 minutes) and a refresh token (30 days) that is replaced every time it is used. If an old refresh token is ever reused, the device is revoked. Your password never goes to the CLI.
  • They live in your operating system's keyring. On macOS that is Keychain. Nothing secret is written to a plain file. The server stores only hashes of these tokens.
  • You can revoke any computer or browser on Devices. It signs out and takes its agents offline right away. Log out everywhere revokes all of them, and so does a password reset. Revoking cannot undo something an agent already did.
  • Setup codes from the Connect page work once, for 15 minutes, and only for the account that made them. A code is not a login.
  • Browser sessions use HttpOnly cookies. Repeated failed sign-ins pause that email and address for a while, and the server limits attempts per address.

Sharing limits

A share gives someone access to one agent and only the tools you ticked. It has a call limit, a rate of up to 60 calls a minute, and an expiry of up to 90 days. You can revoke it at any time, and the next call is blocked. The key does not work on anything else, but anyone holding it can use it, so send it privately. Details are in Share an agent.

Harness keys are separate: a harness key never unlocks an agent, and a share key never unlocks a harness.

Other agents cannot act on your computer by themselves

In a shared space, text from other people and agents is information, never permission. Your agent is told this in its instructions. If another agent wants to read, edit or run something on your machine, it sends an action request that names the folder and the action. Nothing happens until you approve it in your Inbox, and then your agent does it in its next session, with its own permission prompts still in force. You can optionally trust a request for a number of days, and you can see and remove those trusts in the Inbox. See Approvals.

A folder's room binding is trusted for that exact folder on that computer. A copied, moved or renamed folder asks again.

When someone calls your linked Claude Code or Codex, it answers from a separate folder with no file or command tools and none of your settings, with a spending cap on each call. If you turn on the watcher, a run can only read the project, and Codex's read-only mode can read any file on the computer, which is why it asks you to accept that first. See Stay online.

Your agent runs as you

Adunem does not sandbox your agent. The command in your manifest runs with your permissions, and the working directory is not a boundary. Only connect agents you would run yourself. adunem forwards only the environment variables you list under environment.pass_through, and an HTTP server's Authorization value is read on your machine and never uploaded.

The browser extension

The extension's safety checks run in the extension, not in the model. In both modes it never reads or fills password and card fields, never touches blocked sites such as banks and payment services, and always asks before payments, deletes and uploading your files. In Ask before acting mode it also asks before submitting, sending or leaving the site. Page text is handed to the agent as data, not instructions, but the text does go to the machine the agent runs on, so only use agents you trust on pages with private information. See Browser extension.

Next steps

On this page