Architecture of Skitflow

Understanding the architecture of Skitflow is crucial for both deploying and scaling applications. The diagram below shows the current control plane and runtime flow, including the AI Gateway layer that serves the AI Assistant and other AI-powered actions.

Enterprise architecture

Skitflow control plane, workloads, AI, and data fabric

A layered view of the public ingress, Docker Swarm runtime, governed AI access, observability store, and replicated storage.

Self-hosted or managed

Ingress

Internet

Users, APIs, webhooks, agents

TLS + routing

Traefik edge

Ingress, domains, certificates

Traffic enters once, then fans out to the Skitflow dashboard or customer workloads through the same governed edge.

Docker Swarm runtime

Skitflow and user workloads run as services across one or more Swarm nodes.

Skitflow App

Dashboard, API, orchestration

Customer workloads

Apps, Compose stacks, services

AI Assistant

Analysis, diagnostics, proposals

Operational guardrails

RBAC, confirmations, audit trail

Data and observability plane

Core state, queues, and telemetry are separated into explicit backing services.

Postgres

State, config, audit

Redis

Queue and cache

ClickHouse

Logs, traces, metrics

AI governance

Model access stays centralized for built-in and external agent workflows.

AI Gateway

Routing, API keys, usage

AI providers

BYOK or platform models

Storage fabric

LINSTOR replicated volumes

DRBD-backed replicas keep workload data available across Swarm nodes without turning storage into a mystery box under the diagram.

Node A

Replica 1

Node B

Replica 2

Node C

Replica 3

Core components

When you install Skitflow, it sets up the core services that run the platform:

• Next.js application — the frontend, dashboard, API, and control plane logic. Server-side rendering keeps the UI and backend in one cohesive app.
• PostgreSQL — the primary database for organizations, projects, environments, applications, deployments, provider settings, AI routing, and audit data.
• Redis — deployment queue and cache. It serializes work so deployments do not overload the server.
• Traefik — reverse proxy and load balancer. It routes public traffic to both the Skitflow dashboard and the applications you deploy.

Docker Swarm by default

Skitflow runs as a Docker Swarm cluster by default. Start on a single node and add more whenever you need to: Swarm spreads services across nodes and reschedules them if one goes down, without changing how you deploy.

Replicated storage with LINSTOR

When you run more than one node, Skitflow can replicate volumes across them with LINSTOR, backed by DRBD. A service's data is kept in sync on another node ahead of time, so if a node fails, Swarm can reschedule the workload onto a node that already holds an up-to-date replica. See LINSTOR Distributed Storage.

Observability with ClickHouse

Runtime logs, distributed traces, route metrics, and resource metrics are collected by telemetry agents and stored in ClickHouse, a columnar engine built for high-volume telemetry. Human dashboards read from ClickHouse, and the AI Assistant can use the same operational context for diagnostics and proposals. See Observability.

AI Gateway

The AI Gateway is the routing and governance layer for Skitflow AI features. When enabled, Skitflow routes AI calls through a LiteLLM-compatible gateway instead of calling each provider directly from every feature.

The gateway provides:

• Provider routing — route the organization default, Assistant, and Ask AI features to different providers or to the platform default model.
• BYOK and platform models — tenants can use their own provider keys, while managed deployments can provide a platform fallback model.
• Per-organization attribution — gateway traffic uses organization-scoped virtual keys so usage can be attributed and governed.
• Usage and governance — operators can review usage and enforce budgets without exposing provider secrets.

If the gateway is disabled, Skitflow falls back to the direct provider configuration path. If the gateway is enabled but unavailable, AI features fail cleanly instead of hanging on a dead upstream.

AI Assistant

The AI Assistant is the product-facing assistant, not a separate "AI DevOps Agent" anymore. It uses the AI Gateway for model routing and can work with:

• Docker Compose generation and deployment flows.
• Compose review and security hardening.
• Application, compose, database, backup, domain, and environment diagnostics.
• Proposal-based operations that require explicit confirmation before mutation.
• Observability and scoped operational context when available.

Together, these layers let Skitflow deploy applications, operate infrastructure, and route AI work through a controlled gateway rather than scattering provider calls across the product.