Security model
How AgentGarm is built to be trusted: least privilege, hardware-bound approvals, signed everything.
The agent is the adversary
An AI agent runs as you. It can read what you can read, start what you can start, and stop or impersonate any process that runs as you. AgentGarm is designed on that assumption: nothing that runs as you is trusted to decide, and nothing local counts as authenticated.
In a root install
The strongest guarantees need an administrator to install AgentGarm as root:
- The policy, the keys and the audit log are held by
agentgarmd, a process the agent cannot write to. - An approval is a signature from a presence key bound to hardware, which
agentgarmdverifies over the exact action: on a Mac, the Secure Enclave with Touch ID. FIDO2 security keys, Windows Hello and passkeys are built preview.
In no-root mode everything runs as you. AgentGarm still records and decides, but an agent could interfere with it, so a no-root session is rated lower and says so.
Least privilege
The LLM proxy, which reads model traffic, and the launcher, which runs your code, run as you and never as root. Loopback services check the Host and Origin of every request. The session service keeps no long-term secret on disk.
Key custody
The node's key is held in the macOS Keychain, or in a file only its owner can read, and every machine reports its key custody, so a software-held key is visible as such. Hardware-held custody, in the Secure Enclave or a TPM, is built preview.
AgentGarm never routes, stores or reuses subscription credentials such as a Claude.ai session.
Across a fleet
In Central Lite preview, an admin signs in with a passkey, and every change in the console is confirmed with a fresh passkey assertion over that exact change. Each admin has a role, so an auditor can read and export evidence but change nothing else, and people your identity provider removes lose access through SCIM.
A fleet kill order preview needs signatures from several approvers' keys, and each machine verifies them before it stops anything. With the organisation's root held offline by two or more people, no single account, not even Central's own, can stop a fleet alone; under the single software root a small install starts with, whoever controls Central's server could. An incident exports as an evidence bundle of signed audit records that can be checked away from Central preview.
Central's keys can be wrapped by a hardware security module over PKCS#11 preview.

Your data stays on the machine
Prompt bodies stay on the machine. The proxy inspects model traffic locally; nothing is sent to us, and this site collects nothing about you.
Signed releases
Releases are code-signed and notarised; a platform whose signing is not yet in place ships only as a marked preview, and the download page says which. Every release carries SHA-256 checksums and a minisign signature over them: verify a download.
An audit log you can check
Decisions, approvals and alerts are appended to a hash-chained log, and agentgarm verifies the chain, so an edited or removed entry inside it is detected.
Report a vulnerability
Email support@nannertech.com and start the subject with [security]. Please say what is affected and how to reproduce it, and do not open a public issue. We coordinate disclosure with you.