Agents that clock in for the whole team, not just one chat session
CubePlex gives each team a single long-lived agent with shared memory, skills, and a persistent sandbox — reachable from the web and every IM your team already uses.
What it does
CubePlex is a self-hostable platform for what it calls Workspace Agents: one agent per team, with a stable role, shared knowledge, approved tools, and a working environment that survives across tasks. Instead of every member spinning up their own disposable chatbot, the team shares state — memory, deliverables, credentials, and the sandbox’s working tree all belong to the workspace. The same agent is reachable from the web UI and from Slack, Discord, Teams, Feishu, DingTalk, and WeCom, with each conversation keeping its own participants and execution context.
The interesting bit
The persistent sandbox is the quiet differentiator: files, installed packages, and the working tree survive restarts, so the agent resumes at the same “work site” rather than starting from a blank slate every task. The agent runtime itself is built on the team’s other project, CubeLoop, an async framework with durable checkpoints — CubePlex is largely the governance and workspace layer wrapped around it.
Key highlights
- Layered memory: personal preferences, workspace facts, and org policies recalled at the right scope
- Skills from built-ins, org-provided packages, or remote registries like skills.sh
- Governed MCP tools: admins curate a connector catalog; credentials can be org-, workspace-, or user-scoped, encrypted at rest
- Automation via cron, interval, one-shot schedules, and webhook triggers
- Deploys on Docker Compose for a single host or Helm for Kubernetes, with the same images in both modes
Caveats
- It’s a full platform — FastAPI backend, Next.js frontend, sandboxes, IM bridges — so expect real operational weight, not a drop-in script.
- External model providers, MCP servers, and IM platforms sit outside CubePlex’s trust boundary by design; that’s a deliberate constraint, not a bug.
- The README is a feature tour; depth lives in the docs site, so evaluating fit requires a second stop.
Verdict
Worth a look if your team wants a shared, self-hosted agent with real governance and persistent state instead of per-person chat sessions. Solo tinkerers after a quick local agent will find it heavier than they need.
Frequently asked
- What is cubeplexai/cubeplex?
- CubePlex gives each team a single long-lived agent with shared memory, skills, and a persistent sandbox — reachable from the web and every IM your team already uses.
- Is cubeplex open source?
- Yes — cubeplexai/cubeplex is an open-source project tracked on heatdrop.
- What language is cubeplex written in?
- cubeplexai/cubeplex is primarily written in Python.
- How popular is cubeplex?
- cubeplexai/cubeplex has 585 stars on GitHub.
- Where can I find cubeplex?
- cubeplexai/cubeplex is on GitHub at https://github.com/cubeplexai/cubeplex.