Open-source multi-tenant agent runtime for production.
The agent is the tip of the iceberg. Below the waterline: tenant isolation, per-user cost controls, security, billing, deployment. The other 80% that every developer builds by hand. We built it, packaged it, and open-sourced it.
Multi-tenant isolation. Architectural boundaries between tenants. Customer A's data physically cannot reach Customer B's runtime.
Per-user spend controls. Cap what any user costs you in real time, across every runtime boundary – app code, agent harness, generated code in sandboxes.
Dual-sandbox security. Agent-generated code runs with no network, no credentials, no env vars. Privileged operations run in a separate sandbox. Two physical walls, not one.
Channel-based agent protocol. Named channels replace JSON tool-calling. No escaping, no mangled code. Production agents run reliably on Haiku.
8 tools for any system. Generic ontology-aware tools discover connected systems at runtime. No per-integration tool definitions. No retraining.
Deploy from Git. Hot reload. YAML config. Runs on a single server. Kubernetes optional.
Install the CLI and stand up a local platform:
pipx install kdcube-cli && kdcube-setupYou built agents in LangGraph, CrewAI, Claude Agent SDK, or raw API calls. They work. Now you need to run them for users who aren't you.
| DIY stack | + LangSmith | KDCube | |
|---|---|---|---|
| Agent loop | Your framework | Your framework | ReactV3 or bring your own |
| Observability | ✗ | ✓ $39/seat/mo | ✓ |
| Deployment | ✗ | Cloud only | ✓ Git, hot reload |
| Tenant isolation | ✗ | ✗ | ✓ Architectural |
| Cost controls | ✗ | ✗ | ✓ Per-user |
| Billing | ✗ | ✗ | ✓ Cross-runtime |
| Security | ✗ | ✗ | ✓ Dual sandbox |
One architect. AI agents as developers. Thousands of corrections. Continuous supervision. Built with agents so you can build with agents.
MIT
- kdcube — the core platform: backend, CLI, SDK, and chat UI. Production runtime for multi-tenant AI agents: tenant isolation, spend controls, sandboxed execution, deployment. Python · MIT.
- claude-plugins — a Claude Code plugins marketplace for authoring and operating KDCube apps. Shell · MIT.
Install the CLI and stand up a local platform:
pipx install kdcube-cli && kdcube-setup- Docs & site — kdcube.tech
- Core platform — github.com/kdcube/kdcube
- Live demo — kdcube.tech/demo.html
- Releases — github.com/kdcube/kdcube-ai-app/releases
- CLI on PyPI — pypi.org/project/kdcube-cli
- Container images — hub.docker.com/u/kdcube
If KDCube fits how you build, star kdcube to follow along.
From prototype to production, on infrastructure you own. Questions or partnership inquiries: info@kdcube.tech · Released under the MIT License.