System Overview
Skyflo is built from four main components: the Engine (Python, FastAPI, LangGraph), the MCP Server (FastMCP, typed tools), the Command Center (Next.js UI), and the Kubernetes Controller (Go operator). Each runs in your cluster and communicates over internal APIs.
Component Flow
User → Nginx Proxy → Command Center (chat UI) / Engine (FastAPI, SSE streaming) → MCP Server (tool execution) → Kubernetes cluster
An Nginx reverse proxy sidecar runs alongside the Command Center (ui service). It routes /api/v1/* requests to the Engine (port 8080) and all other traffic to the Command Center (port 3000). SSE streaming is supported with buffering disabled. The Engine streams responses via Server-Sent Events (SSE) and invokes the MCP Server (port 8888) for tool execution. The MCP Server runs kubectl and helm as async subprocesses and talks to Jenkins via httpx. The Kubernetes Controller manages Skyflo's own lifecycle via a CRD (SkyfloAI custom resource).
Internal Communication
- Engine ↔ MCP Server: Streamable HTTP transport. Tool calls and results flow through typed MCP protocol.
- Engine internal: Redis pub/sub for streaming events, PostgreSQL for persistence (conversations, messages, token usage).
- Controller: Watches the SkyfloAI CRD and reconciles deployments, scaling, and configuration.
Tech Stack
| Layer | Technologies |
|---|---|
| Backend | Python 3.11+, FastAPI, LangGraph, LiteLLM, Tortoise ORM, Redis, PostgreSQL |
| MCP | FastMCP, httpx, Pydantic, Streamable HTTP |
| Frontend | Next.js 14, TypeScript, Tailwind CSS, Radix UI, Zustand |
| Infrastructure | Kubernetes, Argo Rollouts, Helm |
| Communication | SSE, Redis pub/sub |
| Security | JWT + refresh token rotation, RBAC, HttpOnly cookies |
Component Details
- Engine: LangGraph workflow, LLM orchestration, approval gates, SSE streaming
- MCP Server: Typed tools for kubectl, Helm, Argo Rollouts, Jenkins
- Command Center: Chat UI, real-time streaming, approval UI
- Kubernetes Controller: CRD-based deployment and lifecycle management