zrok is a self-hostable, secure internet-sharing and tunneling platform (ngrok-like) built on OpenZiti zero-trust networking. In production it runs as a distributed Go system: a central controller (REST control plane managing accounts, environments, shares, accesses, namespaces, organizations, frontends, limit classes, and agent enrollment), public and dynamic proxy frontends that terminate internet traffic and route it through the Ziti overlay to private shares, a Cobra CLI (zrok2), a background agent daemon that spawns and supervises share/access subprocesses via gRPC over a unix socket, two React UIs (admin console and agent console), and three client SDKs (Go, Python, Node.js). Share backend modes include HTTP proxy, static web/file, TCP, UDP, SOCKS5, Caddy, and WebDAV drive sync. Persistence is PostgreSQL or SQLite with parallel embedded migration sets; RabbitMQ/AMQP carries metrics and dynamic-proxy mapping updates; InfluxDB stores usage telemetry used for bandwidth-limit enforcement and sparkline queries; SMTP handles verification/password-reset/limit-warning email; OAuth providers (GitHub, Google, OIDC) authenticate public frontend visitors. The architecture is a multi-process control-plane/data-plane split: the controller automates OpenZiti edge resources (identities, services, service policies, configs) through a ResourceManager abstraction with a compensation ledger for cleanup on partial failure, while frontend endpoints and share processes communicate through the Ziti overlay rather than direct TCP. The metrics subsystem ingests OpenZiti fabric.usage events via websocket, validates/enriches them with a retry budget and ACK/NACK semantics, then sinks them to AMQP and InfluxDB. Bandwidth and resource-count limit classes are enforced against InfluxDB reads with deadline handling and relax cycles. The codebase is notably engineering-mature for an open-source tunneling product: design journals, extensive unit/integration tests (including a fake Ziti server for automation tests), pinned codegen tooling, nfpm/goreleaser/Docker packaging, and a repair-store command for orphaned-data cleanup. Weaknesses temper the otherwise strong engineering: a massive generated-code surface (hundreds of go-swagger and OpenAPI Generator files checked in, with duplicated TypeScript model trees across ui/ and sdk/nodejs), mixed logging frameworks (log/slog alongside leftover logrus/pfxlog), a brittle Python SDK workaround parsing an unsupported content-type error body, duplicated cookie/JWT logic between publicProxy and dynamicProxy, and costly schema migrations that rewrite entire SQLite tables on every backend-mode addition. It is not a data-research repository; its data components (metrics pipelines, InfluxDB readers, usage journals) are operational infrastructure for an application.