Vicoa:手机上发任务,电脑同时跑八个 Agent——ADE 全平台开源
Vicoa: Send Tasks from Your Phone, Run Eight Agents in Parallel on Desktop — ADE Goes Fully Open Source
一个被忽视的需求
你有三个任务要跑:一个给 Claude Code 重构模块,一个给 Codex 写测试,一个给 OpenCode 做文档。现在你需要打开三个终端,切换三个目录,分别跟三个 Agent 说话,还要自己保证它们不会踩同一个 Git 分支。
Vicoa 解决的正是这个问题。它的定义很简单:
ADE — Agent Development Environment。 像 IDE 管代码一样,统一管理一支 Coding Agent 团队。
全平台开源刚刚宣布:CLI/守护进程 + Web 前端 + Electron 桌面端 + Flutter 移动端,一套代码,任意设备部署和控制。
支持哪八个 Agent
| Agent | 集成方式 |
|---|---|
| Claude Code | 原生集成 |
| OpenAI Codex | 原生集成 |
| OpenCode | ACP 协议 |
| Google Gemini | ACP 协议 |
| Cursor | ACP 协议 |
| GitHub Copilot | ACP 协议 |
| Kimi | ACP 协议 |
| Hermes | ACP 协议 |
Claude Code 和 Codex 有专门的原生集成;其余通过 ACP(Agent Client Protocol) 接入——这是一个开放协议,意味着以后可以扩展到更多 Agent,而不需要修改 Vicoa 核心。
Worktree 并行:每个 Agent 独立分支
这是 Vicoa 在工程上最关键的设计:每个 Agent 运行在自己独立的 Git Worktree 上。
普通方案是多个 Agent 共享一个工作目录——它们会互相覆盖对方的修改,或者争同一个分支锁。Vicoa 给每个 Agent 分配一个 Worktree,天然隔离,互不干扰,同一个项目可以真正地并行开工。
本地守护进程负责 spawn 每个 Agent 进程,通过 FastAPI + WebSocket 把状态推到任意客户端。你在手机上看到的实时进度,和桌面端是同一个状态流。
移动端:不在电脑旁边也能继续
Vicoa 的 iOS 和 Android 原生 App(Flutter 实现)不是桌面端的简化版——它有几个专门为移动场景设计的功能:
- 推送通知:Agent 任务完成、出错、需要确认——不用盯着屏幕
- 语音输入(Dictation):对着手机说需求,直接发给 Agent
- Git Diff 查看:Agent 改完了什么,直接在手机上 review
睡前把任务发出去,早上起来看结果——这是 Vicoa 想支持的工作模式。
任务面板与定时任务
Vicoa 有一个完整的任务管理层:
- 任务面板(Task Board):类似看板,追踪各 Agent 的任务状态
- 自动化(Automations):cron 定时任务,比如每晚 23:00 让 Codex 跑测试套件,或者每周一早上让 Claude Code 生成周报草稿
- Skills 管理:集中管理各 Agent 可用的 Skill,不需要每台机器单独配置
架构
本地守护进程(Python FastAPI)
├── spawn Claude Code 进程(Worktree A)
├── spawn Codex 进程(Worktree B)
└── spawn OpenCode 进程(Worktree C)
↓ WebSocket
┌──────────────────────────────────┐
│ Web(Next.js 15 + React 19) │
│ Desktop(Electron) │
│ Mobile(Flutter iOS/Android) │
└──────────────────────────────────┘
↓ PostgreSQL
持久化:任务/会话/Agent 状态/Automation 规则
守护进程是核心——它是跑在你本地机器(或 VPS)上的那个进程,管理所有 Agent 子进程。客户端只是界面,可以是浏览器、Electron 窗口、或者手机 App,都连接到同一个守护进程。
安装
CLI(推荐入口)
# Node
npm i -g @vicoa/cli
# 或 Python
pip install vicoa
自部署(Docker Compose)
git clone https://github.com/vicoa-ai/vicoa.git
cd vicoa
docker compose up -d
Docker Compose 包含守护进程 + PostgreSQL + Web 前端,一条命令拉起完整服务。自部署意味着你的 Agent session、任务历史和 API Key 全部留在自己的机器上。
谁需要 Vicoa
同时用多个 Agent 的开发者:你已经在 Claude Code、Codex、Cursor 之间切换——Vicoa 把这些工具统一到一个控制面板里,不再需要手动管理多个终端窗口和分支。
远程/外出工作:任务跑起来之后离开电脑,手机上继续监控和交互,Agent 完成了推通知告诉你。不需要保持 SSH 连接或者 VPN 隧道。
团队 Agent 协作:自部署到内网服务器,团队成员通过 Web 界面共享同一个 Agent 环境,不用每个人各自配置一套本地环境。
Agent 工作流研究:Vicoa 的 Worktree 并行 + ACP 协议栈是一个研究多 Agent 协同的实验平台,可以直接在上面测试不同 Agent 的分工策略。
当前状态
刚刚宣布全平台开源(1 ⭐,初期)。CLI/守护进程、Web 前端、Electron 桌面端、Flutter 移动端均已开源。Claude Code 和 Codex 原生集成就绪;其余 Agent 通过 ACP 接入。自部署文档和 Docker Compose 配置随仓库一起发布。
早期项目,API 和协议有可能变动,但整体架构已经稳定。
总结
Vicoa 做的事情是把”使用 AI Coding Agent”从单个终端操作,升级为一个可以跨设备、多 Agent 并行的完整工作环境。关键设计决策是三个:Worktree 隔离保证并行不冲突;本地守护进程 + 多客户端保证随处可用;ACP 开放协议保证 Agent 可扩展。如果你已经在多个 Agent 之间来回切换,Vicoa 提供了一个统一的控制层。
GitHub: vicoa-ai/vicoa
安装: npm i -g @vicoa/cli 或 pip install vicoa
自部署: docker compose up -d
Vicoa: Send Tasks from Your Phone, Run Eight Agents in Parallel on Desktop
You have three tasks to run: one for Claude Code to refactor a module, one for Codex to write tests, one for OpenCode to generate docs. Today that means three terminals, three directories, three separate conversations — and manually ensuring they don’t collide on the same Git branch.
Vicoa is built for this problem. Its definition:
ADE — Agent Development Environment. Manage a team of coding agents the way an IDE manages code.
The full-platform open-source release was just announced: CLI/daemon + Web + Electron desktop + Flutter mobile, one stack, deployable and controllable from any device.
Eight Supported Agents
| Agent | Integration |
|---|---|
| Claude Code | Native |
| OpenAI Codex | Native |
| OpenCode | ACP protocol |
| Google Gemini | ACP protocol |
| Cursor | ACP protocol |
| GitHub Copilot | ACP protocol |
| Kimi | ACP protocol |
| Hermes | ACP protocol |
Claude Code and Codex have dedicated native integrations. The rest connect via ACP (Agent Client Protocol) — an open protocol that makes future agents addable without changes to Vicoa’s core.
Worktree Parallelism: Every Agent on Its Own Branch
This is the key engineering decision. Each agent runs in its own isolated Git worktree. The typical alternative — multiple agents sharing a working directory — means overwrites and branch collisions. Vicoa gives each agent its own worktree: naturally isolated, genuinely parallel, no locks to manage.
A local daemon spawns each agent subprocess and pushes state via FastAPI + WebSocket to any connected client. The real-time progress you see on your phone is the same state stream as the desktop.
Mobile: Keep Going When You’re Away from the Desk
Vicoa’s native iOS and Android apps (Flutter) aren’t simplified desktop ports — they have features designed specifically for mobile:
- Push notifications: task completed, errored, needs input — no need to watch a screen
- Dictation: speak your requirements, send directly to the agent
- Git diff review: see what the agent changed, on your phone
Send tasks before sleep, review results in the morning — that’s the workflow Vicoa is designed for.
Task Board and Automations
Vicoa has a complete task management layer:
- Task Board: kanban-style tracking of all agent task states
- Automations: cron schedules — e.g., run the test suite every night at 23:00, or generate a weekly draft every Monday morning
- Skills management: centrally manage which skills are available to each agent, no per-machine setup needed
Architecture
Local daemon (Python FastAPI)
├── spawn Claude Code (Worktree A)
├── spawn Codex (Worktree B)
└── spawn OpenCode (Worktree C)
↓ WebSocket
┌───────────────────────────────┐
│ Web (Next.js 15 + React 19) │
│ Desktop (Electron) │
│ Mobile (Flutter iOS/Android) │
└───────────────────────────────┘
↓ PostgreSQL
Persist: tasks / sessions / agent state / automation rules
The daemon runs on your local machine or a VPS and manages all agent subprocesses. Clients — browser, Electron window, phone app — are just views connected to the same daemon.
Install
# Node
npm i -g @vicoa/cli
# Python
pip install vicoa
Self-Host
git clone https://github.com/vicoa-ai/vicoa.git
cd vicoa
docker compose up -d
Docker Compose brings up the daemon + PostgreSQL + web frontend in one command. Self-hosting means your agent sessions, task history, and API keys stay on your own infrastructure.
Who It’s For
Developers already switching between multiple agents: Claude Code, Codex, Cursor — Vicoa unifies them in one control plane, no more managing separate terminal windows and branches.
Remote/away-from-desk work: Start tasks and walk away. Monitor on your phone, get notified when agents finish. No need to keep an SSH connection alive.
Team agent collaboration: Self-host on an internal server, share one agent environment across a team via the web UI — no per-person local setup.
Multi-agent workflow research: Vicoa’s worktree parallelism + ACP protocol stack is a ready-made experimental platform for testing agent division-of-labor strategies.
Current Status
Just announced full open-source release (early, ⭐1). CLI/daemon, web frontend, Electron desktop, and Flutter mobile are all open-sourced. Claude Code and Codex native integrations are ready; other agents connect via ACP. Self-hosting docs and Docker Compose ship with the repo.
Early-stage project — APIs and protocol may evolve — but the architecture is solid.
Summary
Vicoa upgrades “using AI coding agents” from a single-terminal operation to a multi-device, multi-agent parallel environment. Three key design decisions: worktree isolation for collision-free parallelism; local daemon + multi-client for ubiquitous access; ACP open protocol for extensible agent support. If you’re already switching between multiple agents, Vicoa provides a unified control layer.
GitHub: vicoa-ai/vicoa
Install: npm i -g @vicoa/cli or pip install vicoa
Self-host: docker compose up -d
关于本站 · 免责声明
🍄 Mushroom Research Blog 是非营利、免费公开的个人科技观察博客与公众号 XStack18,不接受商业合作、不代表任何企业或机构立场,也不谋求商业利益。我们以个人视角客观中立地记录和分析 AI、Web3 等领域的最新模型发布与技术动态——不止转述新闻标题或二手信息,而是给出有独立思考的深入分析,希望帮更多人获得有价值的一手科技认知。
⚠️ 文中介绍的开源代码与模型,仅供学习交流与技术借鉴。它们大多仍处于早期阶段,有待进一步研究和验证,请勿直接用于工作或生产环境;如需采用,请先自行充分测试,并核实其许可证与安全性。
Open-source code and models featured here are shared for learning and reference only. Most are early-stage and still need further study and verification — please don't use them directly in your work or in production. Test them thoroughly and check their licenses and security first.
- 本站文章均为作者基于公开信息的个人研究与观点整理,不代表文中提及的任何公司、产品、模型的官方立场,未与其构成商业关联或合作关系。
- 科技行业信息更新极快,我们尽力保证内容准确、及时,但不对完整性、实时性做绝对保证,具体请以相关企业/项目官方公告为准。
- 文中引用的第三方商标、产品名称、图片、数据等版权归原权利人所有,我们会尽量注明来源;如你认为存在版权疑问或侵权,请通过下方邮箱联系我们,收到通知后会尽快核实处理(更正、加注来源或删除)。
- 文章内容仅为技术科普与个人观点,不构成投资、法律或其他专业建议,据此进行任何决策的后果需自行判断和承担。
📮 侵权 / 勘误 / 合作咨询:hello@mushroom.cv
💬 评论与讨论
使用 GitHub 账号登录后发表评论