Authsia
Runtime security for AI coding agents
Agents can run commands. They don’t inherit your secrets until you say so.
Local Keychain. Scoped JIT. Revoke anytime.
Access Center
Agent secret request
- Caller
- Codex
- Wants
- Project/API
- Scope
- exec · this workspace
- Status
- Waiting Approved · 14:37
Injected into child process only
Parent shell stays clean. Masked on mediated output.
Agents inherit your authority
A coding agent runs with your permissions. It can launch tools, touch cloud CLIs, and act on your machine. A vault that only protects secrets at rest does not decide whether that agent should receive a credential at runtime.
Authsia closes that gap on your Mac — locally, with a human in the loop.
Every secret use is a grant
Access Center shows who is asking, what scope they want, and how long access lasts. Approve, deny, or revoke without changing project files.
Caller
Know who is asking
Agent and tooling identity stays visible for the life of the grant.
Scope · TTL
Bound the blast radius
Time-boxed, workspace-aware access — not a standing secret in the parent shell.
Revoke
Cut it off anytime
End the grant immediately when the task is done or something feels wrong.
Secrets leave the vault at the last mile
Authsia resolves commit-safe refs and injects credentials into approved child processes. The agent-visible parent shell stays clean, and known secret values can be masked on mediated output.
Local audit history records what was released, who approved it, and what ran under that grant — without sending telemetry away.
# run with scoped secrets
authsia workspace run -- npm test
# review what was released
authsia audit listInstall on your Mac
Launch Authsia once, enable CLI Access, then let workspace commands carry the security boundary for terminals and agents.
# app + CLI
brew install --cask james-liang-cs/authsia/authsia
# first check
authsia status