Security model
Privileged operations stay in the main process, the renderer gets only a curated IPC surface, and risky permission states remain visible rather than silently escalated.
AgentDock's security posture is mostly a consequence of its process model rather than a separate bolted-on layer.
Structural boundaries
- Context isolation — the renderer has no direct Node.js or Electron API access; everything privileged goes through the preload bridge's curated, typed surface.
- Local data only — session, workspace, and settings data lives in a local SQLite database with no remote backend to secure or leak from. See Security and privacy.
- Provider credential separation — AgentDock never stores or handles authentication for any agent CLI; that stays entirely between you and each provider.
No silent escalation
Permission modes that skip prompts entirely (bypassPermissions, danger-full-access, bypass) remain visible in the UI rather than being hidden defaults — and AIgency's own bypass run-level policy requires a one-time acknowledgment before it can be used. See Recommended practices.
Compare Mode hardening
Compare Mode forces every pane into a safe, read-only-equivalent mode, and (in AgentDock Pro) auto-declines any prompt from an agent attempting to escalate out of that forced mode mid-comparison — see Read-only isolation.
No sandbox-disable workaround
AgentDock deliberately does not work around a broken Linux Electron sandbox helper by launching with --no-sandbox — see Linux sandbox issues for the correct fix.