Local-First AI Security.

For the API proxy, the recommended default is TLS interception via a local CA, which gives full coverage of direct LLM API traffic regardless of how each client is configured, including apps that hardcode the provider URL. A cert-free mode is also supported, where coverage is limited to traffic routable through patched SDK base-URL configuration; customers who decline the cert accept the narrower visibility. The browser extension needs no certificate on its surface, since it reads content after TLS terminates inside the browser.

EDR sees process-level activity, malware behavior, and post-breach signals. The MagicMirror daemon sees AI-specific activity at the workspaces where AI happens: prompts, tool calls, API requests, file writes, clipboard moves, agent actions. Different layer, different problem. We coexist with major EDR vendors and validate it during the pilot.

The synchronous enforcement hooks wired into coding agents fail open if the daemon is unreachable, so developer workflows aren't broken by an outage. For surfaces where you'd rather classify and log without blocking, observe-only mode is available, delegating enforcement to the hook layer in the agent's execution path.