What KDCube Is

KDCube is a self-hosted platform and SDK for packaging AI applications into deployable bundles and running them in isolated environments. tenant/project is one environment, one environment can host many bundles, and one bundle is one end-to-end application unit that can combine backend logic, APIs, widget UI, main UI, agents, tools, ISO runtime execution, and scheduled work. Many bundles use the built-in ReAct subsystem with tools, skills, continuous conversations, a sources pool, ANNOUNCE, a signals board, and an event timeline. The platform then provides the core services that run that model: Ingress (API gateway, auth, SSE emitter) and Processor/Proc (bundle execution, queue worker, integrations API).

Bundle UI surfaces are served by KDCube the same way bundle APIs and MCP endpoints are served. A bundle can expose widget UI and/or main UI; a frontend may display those surfaces directly or embed them, for example in the KDCube control plane. The iframe is only one display mechanism, not a separate bundle concept.

System at a Glance

System architecture overview diagram
System Architecture Overview Platform architecture showing Ingress and Processor services with client, bundle, and storage interactions User Browser / App SSE + REST SSE/WS REST /bundles/{t}/{p}/{b}/operations/{op} Ingress Auth & Rate Limit Bundle Routing SSE Emitter Task Enqueueing enqueue Redis Queue Task Buffer dequeue Processor (proc) Bundle Loader LangGraph Exec Communicator (async) Operations API (REST) GET/POST /bundles/{t}/{p}/{b}/operations/{op} invoke YOUR BUNDLE Entrypoint Workflow Tools Skills ReAct Agent Widgets / UI Storage · Config · Economics Communicator Firewall Settings spawns ISO Runtime (Docker / Fargate) Supervisor (networked) All bundle tools network · provider Communicator → Redis Pub/Sub → SSE (same path as proc) ↔ Unix socket Executor (no network) LLM-gen code tool_call() → Unix socket ✗ no network ✗ no secrets SSE stream Communicator (async streaming → client) PostgreSQL Conversations Redis Props / Cache S3 / Local FS Artifacts / Files Platform-managed persistence layer

Key Terms

TermDescription
BundlePython package registered with @agentic_workflow. The unit of AI application deployment — contains agent logic, tools, skills, widget UI, main UI, operations, and optional public endpoints.
IngressAPI gateway service: auth, SSE streaming, task enqueueing, rate limiting. Handles all inbound traffic before the bundle sees it.
Processor / ProcQueue worker: executes bundles, hosts the Operations REST API. Loads bundle singletons and calls execute_core() per turn.
TimelineRolling cross-conversation context persisted as artifact:conv.timeline.v1. Ordered blocks (oldest → newest) streamed as SSE events. Grows across turns; compacted when context budget is reached. The turn log is the per-turn portion appended each turn.
ReAct AgentAutonomous timeline-first loop with bounded iterations. Plan is a tool — not a separate component. No coordinator needed.
Tenant / ProjectMulti-tenancy units. All data, config, bundles, and budgets are scoped to tenant + project. Each tenant gets an isolated schema in PostgreSQL.

Supported Providers

🔌
Transports: SSE · Socket.IO · REST  |  LLMs: OpenAI · Anthropic · Gemini · OpenRouter  |  Search: Brave · DuckDuckGo  |  Auth: Cognito · SimpleIDP · Delegated