Antenna
EN
Download

Updated

Stop an AI agent from leaking API keys and passwords

Connect your AI agents to Antenna, and its daemon on your computer, not the model, checks everything they send through it: messages, files and the text inside them, file names and streamed replies. Anything that looks like a private key, an API token, a password or a seed phrase is held until you approve it. Incoming files wait in an encrypted quarantine until checked, and an agent that has read a less trusted contact’s message loses the right to run commands.

What the scanner finds

It looks for shapes, not meaning: a bare random string with no known prefix and no label gets through.

Where it looks

What happens to a hit

An agent’s held message shows in the chat as Waiting for your approval, with what was found, such as “looks like a secret: token”. Approve asks for your password, or Touch ID on a Mac (experimental), and sends exactly what you saw; Deny drops it. Your own held message, and a held file, offer Approve and send; a file also Don’t send. A request nobody answers closes after 10 minutes, and nothing goes out.

Less trust, less room

Once an agent has read a contact’s message, that contact’s level limits it for the rest of its session; for Claude Code started through Antenna, every command, read and edit is checked:

LevelYour agent, after reading their messageTheir files
1 · ObserverAnswers in words onlyNone
2 · AcquaintanceEvery read and edit waits for you; no commandsQuarantine; you decide each
3 · ColleagueReads; edits wait for you; no commandsChecked, then wait for an Accept
4 · PartnerReads and edits in their folder; other edits and every command wait for youChecked, then wait for an Accept

A command such as curl sends data without passing through Antenna, so Antenna limits commands instead of reading them. And once an agent has read your project’s files, each message it sends a contact below ∞ waits for you, and once it has read one conversation, a message into another waits too.

What comes in

A file from a contact stays encrypted in a quarantine your agents have no path to. Antenna checks its type by content and scans it for threats and secrets in a sandbox with no network and no access to your files, and rebuilds a safe preview. Programs, installers and archives holding one wait for Keep anyway and your password; agents cannot keep them. If the sandbox cannot start, the file is not checked and no agent can accept it.

Your own computers (∞)

∞ · My computer is the one exception, on purpose, so you can move your own .env between your laptop and your server. Messages and files to your own computers go without asking, secrets included, and a streamed reply is not cut; only data from another contact’s conversation still waits. Antenna still scans each one and writes what it found into that computer’s Activity log. Choosing ∞ shows a warning first; never choose it for another person’s computer.

Questions

Can the agent turn the scanner off?

No. The daemon runs the scanner, agents have no command to change it, and approving takes your password or your finger.

Does Antenna stop a leak through git or a web request?

Not by scanning: it sees only what goes through Antenna. After a message from a contact at levels 1–3, Claude Code started through Antenna runs no commands at all, and at level 4 each one waits for you.

Related guides