Skip to content
Perfecto23Public

About

AI agent workspace for teams

Resources

Contributing

Stars

0 stars

Watchers

0 watching

Forks

 
 

Repository files navigation

CubePlex × AgentCore

这是基于 cubeplexai/cubeplex 的 fork,上游基线为 f5272e19(产品版本 0.7.2)。本 fork 将 CubePlex 的产品控制面与 Agent 执行环境拆开:员工继续使用原来的网页或 Slack,CubeLoop Agent 和 Git/Shell/文件工具在 AgentCore MicroVM 中按需运行。

员工使用 Web / Slack
  → Kubernetes 上的 CubePlex
      账号、工作区、会话、任务状态、消息收发与持久化数据
  → AgentCore Native Runtime
      CubeLoop Agent + Git/Shell/文件工具
  → 任务回调返回 Backend
      保存检查点、进度、人工确认和文件,再展示给员工

当前 Testing 使用 cubeplex_native_entry_20260913-sWxaCf7pyC Runtime v2,状态 READY。Native 真实网页与 Slack 任务、跨会话追问、下载文件、人工确认、Backend 重启、停止与回收、重复消息和晚到回调均有对应验收记录;本次 cleanup Web readback 为 CLEANUP-NATIVE-OK。关闭本机运维连接后,网页与 Slack 的新任务仍由云端处理;没有进行物理电脑关机实验。

实现与部署

  • Kubernetes 控制面:复用现有单节点 k3s、Backend、Frontend、Postgres、Redis、RustFS 和原生 Slack Socket Mode,入口不依赖本机轮询程序。
  • Native 执行层:自制 ARM64 镜像运行 CubeLoop 与本地工具。模型、数据库和存储主凭据留在 Backend;MicroVM 只有对应任务的短期回调权限和最小 AWS 角色。
  • 状态与恢复:检查点与请求回执在同一 PostgreSQL 事务提交;事件、终态和停止有去重与隔离。等待人工回答时可以回收 VM,回答后用新会话继续同一个 run。
  • 运行边界:Native Runtime v2 是当前唯一产品执行路径。OpenSandbox 继续承载 Browser、Terminal 和文件侧栏;旧 compatibility Runtime、Git slice Runtime、Lambda、S3 状态和专属 Secrets 已退休。

Native Worker 来自 09f272ec4088f72fe709cc89b01d7abd68fe45d3,Backend 来自 5e616137415f7b93d3e86b9514739ba129d9a7c4,镜像 digest 与部署方法见 Native 运行指南。Backend 主容器和 migration 初始化容器须使用同一新版镜像。Frontend 沿用已部署版本。已修复有供应商补丁的镜像问题;仍保留已知未修复的 zlib High,Native 镜像另有 nghttp2 Medium,不能宣称扫描清零。

当前能力边界

原生路径支持公开仓库的 clone、读取和命令执行,以及普通文件的保存、展示和恢复。真实 Git 测试首轮因 Agent 未切换目录而发现 0 个测试,未计为通过;同会话纠正后实际运行了 4 个测试(1 通过、3 失败,测试仓库 main 故意保留缺陷),结果已如实保存。

普通文件快照不保存 .git、隐藏文件、凭据、安装环境或后台进程。Browser、Terminal 和文件侧栏仍使用 OpenSandbox 环境;Native 的实际用户入口是对话和下载卡片。独立 Git slice 只保留历史证据,未接入产品入口;其 Worker ECR 基础镜像仍因 Native Dockerfile 的 FROM 依赖保留。员工私库授权、200 仓库检索、AgentCore Browser 和其余工具迁移继续分批实施。

这是单节点 Testing PoC,节点故障没有 HA 保证。约 $3.89/天是历史基线估计,不是本次重新计价;当前持续计费项包括节点、EBS、EIP、ECR/Native 基础镜像、模型和 AgentCore 按量费用。本阶段没有新增 EC2、EKS、NAT Gateway 或数据库,当前只保留一个 Product KubeconfigSecret。

文档入口

阅读目的 入口
当前 Native 路径、镜像、配置、验收和边界 Native 运行指南
现有 AWS/k3s/Helm/Slack 基础设施与 Native 运维 产品部署指南
当前 Native 产品部署架构与边界 产品部署说明
已退休的独立 Git/Shell 实验与历史证据 Git execution slice
标准上游 Docker Compose / Kubernetes 部署 部署入口
历史窄范围 Slack PoC PoC 指南 · 验证记录

独立 Git slice 的真实修复、测试、commit、受限 push、测试 PR 和新 VM 续聊仅作为历史证据保留。它的 Runtime、Lambda、S3 状态和专属 Secrets 已退休;Worker ECR 中的 sha256:454e290... 仍是 Native Dockerfile 的构建基础,不应随实验资源一并删除。该实验不能替代 Native 产品验收,也不能被当作 Native 已接入 authenticated push/PR 的证据。

上游 CubePlex

以下保留上游的产品介绍、演示和文档入口。这些描述属于完整 CubePlex 产品;本 fork 的 AgentCore 产品接入边界、Testing 操作和当前状态以上方产品部署指南为准, 旧 PoC 仅用于历史证据。

CubePlex

Cloud-native platform for managed agents in team workspaces

CI Docs Website Python 3.12+ Node 20+ Docker | Kubernetes

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 architecture: clients, the application and agent runtime, workspace sandboxes, external services, and persistent infrastructure

The diagram reflects the current application architecture. CubePlex's agent runtime is built on CubeLoop, an async-native agent framework for multi-provider model access, tool execution, streaming, middleware, and durable checkpoints. Workspace sandboxes are isolated execution environments with persistent working state; external model providers, MCP servers, and IM platforms remain outside CubePlex's trust boundary.

Demos

interactive-website-demo.mp4

Build Interactive Website — a full product website generated from a single prompt.

skills-workflow-demo.mp4

Skills Workflow — find a skill, install it and use it to build agentic frontend, end to end.

data-analysis-demo.mp4

Data Analysis — transform raw tabular data into a formatted spreadsheet.

one-page-pdf-demo.mp4

One-Page PDF — turn a one-page PDF into a polished, navigable page.

browser-control-demo.mp4

Browser Control — an agent drives the browser to complete a task autonomously.

Features

Area What you get
Multi-model chat Hosted and custom providers (Anthropic, OpenAI, and more). Attach files, stream replies, switch models mid-conversation.
Skills Packaged agent capabilities — built-in, org-uploaded, or from remote registries (e.g. skills.sh).
Memory Personal, workspace, and org-scoped memory the agent recalls across conversations.
MCP tools Catalog of connectors with static credentials or OAuth; grant tools per workspace.
Workspace sandboxes Per-workspace isolated runtimes with persistent storage — files, packages, and the working tree survive restarts so agents resume the same work site.
Artifacts Versioned deliverables — files, previews, code, images — rendered in the thread.
Automation Scheduled tasks (cron / interval / one-shot) and webhook event triggers.
IM bridges Talk to agents from Slack, Discord, Teams, Feishu, DingTalk, and more.
Team governance Organizations, workspaces, roles, model access policies, and cost tracking.
Deploy anywhere Docker Compose for a single host; Helm for Kubernetes.

Get started

Both modes use the same backend and frontend images. Guides cover image builds, configuration, secrets, and verification.

Develop locally

Prerequisites: Python 3.12+, Node.js 20+, pnpm 10+, and Docker (recommended for local services).

git clone https://github.com/cubeplexai/cubeplex.git
cd cubeplex
make install

# Terminal 1 — API
cd backend && python main.py

# Terminal 2 — web UI
cd frontend && pnpm dev

Backend: http://localhost:8000 · Frontend: http://localhost:3000.

Local setup also needs backend env/config files described in the contribution guide.

Repository layout

backend/    FastAPI API and Cubeloop-based agent runtime
frontend/   Next.js web app and shared TypeScript packages
deploy/     Docker Compose and Kubernetes/Helm assets
docs/       Product docs site and engineering reference
scripts/    Worktree provisioning and dev helpers

Documentation and contributing

About

AI agent workspace for teams

Resources

Contributing

Stars

0 stars

Watchers

0 watching

Forks

Releases

Packages

Contributors

Languages