Bundles — the unit of an AI app
One folder packages backend logic, APIs, widgets, full UI, agents, tools, storage, secrets, and cron. Hot-reloadable; deploy without restarting the runtime.
Bundle docs →What is KDCube
The runtime for your AI mini-apps.
A self-hosted Python runtime. You ship apps as bundles — chat widgets, smart APIs, full UIs, scheduled jobs, MCP servers — and the runtime owns the operational layer beneath them. One management surface, however many apps you ship. Because it runs on your infrastructure, you own the data, the audit perimeter, and the upgrade path.
Per-environment isolation
A tenant/project is a full environment with its own descriptors, secrets, Postgres schema, and Redis. Stage, customer segment, or team — each gets a clean boundary.
Many surfaces, one bundle
Chat widgets, embedded UIs, REST endpoints, iframe apps, MCP servers, and @cron jobs all in one bundle. The default answer to "where does this AI feature live?"
Pluggable agent runtimes
Ship with ReAct v3 (timeline-first) and Claude Code (workspace-bound). Or bring your own inside a bundle — LangGraph, CrewAI, plain Python.
ReAct v3 docs →Built-in economics
Per-tenant, per-customer, per-turn cost accounting. Hard spend caps reserved before each turn. Stripe-integrated wallets and an exportable ledger.
Economics docs →Decision provenance
Every turn is a replayable timeline artifact: sources, plan, tool calls, costs. Replay why an agent did what it did, end-to-end, without scraping logs.
Timeline docs →From prototype to production — on infrastructure you own.
How it works →