It outlives the chat
A key in a prompt is in a log, a sync, and someone else's retention window.
One local vault. Agents in Meta Muse, Cursor, Claude Code, Codex, and VS Code ask for a capability. An on-device Llama 3.2 1B (Edge / WebGPU DirectML) scores the request. You approve. The secret never enters the chat.
A key in a prompt is in a log, a sync, and someone else's retention window.
Rotation starts after the leak. The bill does not wait for your dashboard.
Agents ask all day. A password manager prompt was not designed for that.
AES-256-GCM. Master key in the vault you own. Hard ceilings. Secure even with no model at all.
Default Llama 3.2 1B (Edge / WebGPU DirectML) on this machine. It sees the request, never the key.
After the call, after 60 seconds, or on one revoke. Standing approval is a low read only.
Nothing is stored. These buttons only show which gate fires.
Agent, IDE, alias, scope, reason. That is the whole prompt.
The policy floor. A write under a read ceiling is dead before the model runs.
Twenty-seven surfaces. The build fails if the secret shows up. The sweep.
The agent can ask, list aliases, enroll, check status, and revoke. The snippet.
90% of the code was written by frontier model Muse Wonder 1.0.
Verified defense across all OWASP LLM Top 10 vectors. The matrix.
Built by former crypto wallet bug bounty hunters. Securing keys on-chain since 2025.