Zervanor — Architecture

One event backbone. Zero hardcoded providers.

Any event source. Any business outcome. One platform. Here is the full multi-tenant, Kubernetes-native stack under every ready automation.

Connectors

  • Stateless HTTP calls
  • Lightweight, zero cold start
  • REST / GraphQL / Webhooks
Stripe · Slack · Gmail · GitHub · Notion
Event RouterCloudEvents

Agents

  • Isolated K8s pods
  • Long-running, stateful
  • AI / LLM / Messaging
WhatsApp · AI Agent · Scheduler · Approval / HITL

The full platform stack

Eight layers, one contract. Each layer is independently deployable and observable — and every request ultimately routes through the CloudEvents backbone.

Surfaces
Next.js DashboardVisual BuilderTypeScript SDK + CLI

Visual for operators, code for developers.

Control Plane
FastAPIControllers → Services → RepositoriesJWT + RBAC (4 roles)WebSocket real-time

Auth, org-scoping, and the API every surface speaks to.

Workflow Engine
Prefect42 node typesLive execution monitoring

Matches events by pattern, cron, or webhook and runs the graph.

Event Backbone
CloudEvents 1.0Apache Kafka (Redpanda)24h dedup

The one contract every integration speaks. Highlighted because everything routes through it.

Agent & Connector Runtime
Generic Agentic Runtime (ADR-038)Connector packsTool projection (AG-02)

Connectors call APIs; agents reason in isolated pods. Same packs power both.

Orchestration
Deploy WorkerK8s agent lifecyclePer-tenant namespaces

Containerizes and deploys agents into isolated tenant namespaces.

Data & State
PostgreSQLTimescaleDB (event history)RedisQdrant (vector memory)

Relational state, time-series event history, cache, and agent memory.

Infrastructure
Kubernetes / GKETLS 1.3OpenTelemetry · Prometheus · Grafana

Multi-tenant cluster with encryption, network policies, and full observability.

What the stack gives you

CloudEvents Gateway

Every event source — webhooks, schedules, databases, messaging, APIs — normalized into the CloudEvents 1.0 contract on a Kafka backbone, so templates orchestrate across tools without custom glue.

TypeScript Agent SDK & CLI

Build and ship agents in TypeScript. Scaffold with init, run locally, and deploy with push — pure HTTP, no local Docker required.

Visual Workflow Builder

A visual drag-and-drop builder on a Prefect engine — 42 node types (switch, filter, loop, transform, HTTP, run-agent, connector-action, human-task, and more) plus live execution history.

Schema-as-Contract Connectors

Integrations are data, not code. Each connector is a versioned pack validated against a canonical, API-served schema before it can ship — the catalog scales without touching the core.

Real-Time & Observable

WebSocket live updates with auto-reconnect, plus OpenTelemetry, Prometheus metrics, Grafana dashboards, and a token-free public status page.

Multi-Tenant Security

JWT + RBAC (4 roles), per-tenant Kubernetes namespaces, org-scoped queries, TLS 1.3, and PII redaction — validated by a passed audit (47/47 CI gates).

Built on a modern, open stack

FastAPINext.jsNode.js / TypeScriptKubernetes / GKEPostgreSQLTimescaleDBRedisKafka / RedpandaPrefectQdrantNovuOpenTelemetryPrometheusGrafana

Why this architecture wins

Provider-agnostic by construction

There is zero hardcoded provider logic in the core. The same engine runs WhatsApp, Stripe, and Salesforce — and the next hundred — through the contract. A connector is only "done" with a provider pack, connector pack, full docs bundle, and working test targets: no half-finished, undocumented integrations.

Event-driven execution

Unlike polling-based platforms, Zervanor is event-driven. Workflows and agents wake only when a subscribed event — like stripe.payment.succeeded or github.issue.opened — flows through the Event Router.

Kubernetes-native runtime

Zervanor doesn't just call your APIs — it hosts your agents. Each agent is containerized and deployed into a multi-tenant cluster on GKE, with pod-level isolation and per-organization namespaces. Event history lives on TimescaleDB; observability runs on OpenTelemetry and Prometheus.

New — runtime & models

Run on the platform — or on your own device.

The same agent runs in the managed cloud or on your own machine, draws on a routed mix of cloud and on-device models, and is reachable through a chat interface.

Cloud or your device

The same agent — one manifest, one SDK — runs in a managed cloud pod or on your machine via the Zervanor Desktop app. It connects outbound, so it works behind a firewall.

Smart model routing

One catalog of cloud, platform and on-device models behind named routers with automatic fallback. Choose a model per agent; run inference locally when you want.

Conversations

A persistent, model-routed chat that can drive your connectors and agents as tools — with a confirmation step before anything changes.

Desktop runtime host today; a mobile host is on the roadmap.

See the foundation in action