cubeplexai/cubeplex

CubePlex is a cloud-native platform for managed agents in team workspaces — skills, shared memory, MCP tools, persistent sandboxes, governed access, and self-hosted deploy on Docker Compose or Kubernetes.

CubePlex – Cloud‑native platform for managed AI agents

What it is – CubePlex is an open‑source, cloud‑native application that lets teams create, run, and govern AI “agents” (chat‑style assistants) inside isolated workspaces. It bundles a FastAPI backend (built on the CubeLoop async agent framework) with a Next.js web UI, and can be deployed via Docker Compose or Helm/Kubernetes.

Core ideas

  • Managed agents – agents can call any LLM provider (OpenAI, Anthropic, etc.), run tools, and keep checkpoints.
  • Team workspaces – each workspace gets its own sandbox with persistent storage so an agent can pick up where it left off.
  • Skills & memory – reusable capability packages ("skills") and scoped memory (personal, workspace, org) that survive across chats.
  • MCP tools & IM bridges – connectors to external services (OAuth, static credentials) and chat platforms (Slack, Discord, Teams, etc.) so agents can be invoked from the tools teams already use.
  • Governance – orgs, roles, model‑access policies, and cost tracking built in.

Key features (as listed in the README)

Area What you get
Multi‑model chat Switch between hosted/custom LLM providers, attach files, stream replies, change model mid‑conversation
Skills Packaged agent capabilities – built‑in, org‑uploaded, or fetched from remote registries
Memory Personal, workspace, and org‑scoped memory that agents recall across sessions
MCP tools Catalog of connectors with static credentials or OAuth, grantable per workspace
Workspace sandboxes Isolated runtimes with persistent storage (files, packages, working tree)
Artifacts Versioned deliverables (files, previews, code, images) rendered in the chat thread
Automation Cron/interval/one‑shot tasks and webhook triggers
IM bridges Slack, Discord, Teams, Feishu, DingTalk, etc.
Team governance Organizations, workspaces, roles, model‑access policies, cost tracking
Deploy anywhere Docker Compose (single host) or Helm for Kubernetes

Architecture snapshot

  • Clients – web UI (Next.js) and IM bots connect via HTTP/WebSocket.
  • Application layer – FastAPI server hosts the CubeLoop‑based agent runtime, handling model calls, tool execution, streaming, and durable checkpoints.
  • Workspace sandboxes – per‑workspace containers/pods with persistent volumes that keep files and state across restarts.
  • External services – LLM providers, MCP (model‑connector‑platform) servers, and messaging platforms remain outside the trust boundary; CubePlex only talks to them through configured connectors.

Getting started

  1. Prerequisites – Python 3.12+, Node 20+, pnpm 10+, Docker (optional for local services).
  2. Clone and install:
    git clone https://github.com/cubeplexai/cubeplex.git
cd cubeplex
make install   # sets up backend virtualenv and frontend deps
  3. Run locally:
  4. Deploy to production using the provided Docker Compose file for a single host or the Helm chart for a Kubernetes cluster (links in the README).

Typical use cases demonstrated in the repo

  • Generating a full product website from a single prompt.
  • Installing and using a “skill” to build an agentic frontend.
  • Turning raw tabular data into a formatted spreadsheet.
  • Converting a one‑page PDF into a navigable web page.
  • Autonomous browser control to complete web tasks.

Where to learn more

  • Docs site: https://docs.cubeplex.ai
  • Core concepts: docs/site/docs/getting-started/core-concepts.md
  • Deployment guide: deploy/README.md
  • Contribution guide: CONTRIBUTING.md

Bottom line – CubePlex provides a production‑grade stack for teams that want to run AI agents with persistent state, tool integration, and fine‑grained governance, all deployable on‑premise via containers or Kubernetes.

Related

  • Project
    michael-denyer/pstack-claude

    pstack‑claude is a MIT‑licensed plugin that ships 54 reusable “Agent Skills” (e.g., TDD, CI fixing, PR summarising) for Claude Code and, via a shared skill directory, for Codex, Prime Agent, opencode and Gemini‑CLI. It auto‑injects a poteto‑mode mandate on Claude sessions, provides generated slash‑command stubs for Codex, and includes CI to keep the skill set in sync with the upstream Cursor version.

  • Project
    Socialpranker/deepdive

    Deepdive is an open‑source Claude Code skill that turns a plain research request into a reproducible 13‑phase pipeline. It reframes the question, builds a detailed plan, runs parallel sub‑agents across 29 curated search channels, triangulates and filters evidence, performs adversarial red‑team review, and verifies citations for liveness and faithfulness. The output is a structured folder of markdown files (plan, sources, claims, memo, full report) with every claim linked to its source, making the research auditable and reusable. Installation is a one‑line git clone for Claude Code, and the methodology can be adapted to other LLMs.

  • Project
    modular/mojo-gpu-puzzles

    An interactive, open‑source tutorial that teaches GPU programming in Mojo through hands‑on puzzles, complete with a `pixi`‑based environment, test harnesses, and community support.

  • Project
    ptnghia-j/ChordMiniApp

    An open-source music analysis tool that provides automatic chord recognition, beat tracking, and melody transcription with interactive visualizers.