Self-host CubePlex

CubePlex gives each team a durable AI agent that keeps its role, knowledge, and tools across tasks. The agent runs on CubeLoop, an async framework for multi-provider model access and durable checkpoints. Users reach the agent through a web interface or via bridges to chat platforms. Memory is layered at personal, workspace, and organization levels. Skills package reusable workflows from built-in capabilities or remote registries. MCP connectors are governed through an admin-curated catalog with scoped credentials. Code executes in isolated, persistent sandboxes. Secrets are encrypted at rest. Multiple model providers are supported with mid-conversation switching. Automation runs on schedules or webhooks. Artifacts are versioned and stay in the conversation. Governance covers organizations, workspaces, roles, model policies, and cost tracking. Deployment uses Docker Compose or Helm with the same container images. The stack includes PostgreSQL, Redis, rustfs object storage, OpenSandbox for runtimes, and optional Tempo tracing and docling-serve parsing.