Show HN：自托管企业级 OS OtoDock，基于 Claude Code 与 Codex 构建多部门协同 Agent

OtoDock — 协同 Agent 平台

您企业的大脑，基于 Claude Code 与 Codex 构建，直接在您的 Anthropic 和 OpenAI 订阅上运行。

文档 · 安装 · 功能特性 · 讨论区

自托管 · 原生多租户设计 · Fair Source 协议
支持运行在 Claude Code · Codex · 自有 API Key · 本地模型之上

控制台亮点。两分钟的视频完整展现了一切。

如果 OtoDock 对你有帮助，为本仓库点亮 Star 能让更多人发现它。

---

OtoDock 充当您公司的大脑。您可以创建强大的 Agent，连接企业运行所依赖的工具，按部门协同工作、相互委托任务，并在无人值守时持续自主作业。平台原生采用多租户设计，支持多人协同使用相同的 Agent，并提供四种模式来决定 Agent 的共享机制及其产出落地的位置。您可以在自托管的仪表板上与 Agent 协同，甚至可以为其分配电话线路进行实时通话。所有功能均可在您自有的 Anthropic 和 OpenAI 订阅以及本地模型上运行。

每个 Agent 都拥有专属名称和岗位职责（例如个人助理、系统管理员或市场经理），并完全运行在您自己的服务器上。一个 Agent 由六个模块构成，每一项均支持自定义配置：

• 人设（Persona）。 使用通俗自然的指令告知 Agent 它是谁以及如何开展工作。
• 记忆（Memory）。 Agent 会跨对话记录并维护其专属笔记。
• 工作区（Workspace）。 日常工作成果存放的文件夹。根据 Agent 的模式，可以是个人私有、团队共享，或两者兼备。
• 知识库（Knowledge）。 参考资料随时随地可供调用。
• 技能（Skills）。 Agent 可学习并应用到实际工作中的各项专业技巧。
• 工具（Tools）。 Agent 仅严格调用被明确授权的工具。

共享单个 Agent 的四种模式

每个 Agent 都有特定的模式，决定团队成员如何共享使用：

模式	具体含义
仅个人私有（Personal only）	每位成员拥有专属的独立工作区。
个人优先 + 共享（Personal + shared）	默认为每位成员提供私有工作区，同时配备团队共享工作区。
共享优先 + 个人（Shared + personal）	默认以团队共享工作区承载日常工作，并为每人提供独立的私有工作区。
仅共享（Shared only）	全员共享同一段对话历史和同一个工作区。

为每个人分配合适的角色席位

管理员可将团队成员加入到他们所需要的 Agent 中。

三种平台角色	每个 Agent 上的三种角色
管理员（Admin）：管理并运营整个平台。	主管（Manager）：对该 Agent 拥有完全控制权。
创建者（Creator）：创建新的 Agent。	编辑者（Editor）：可编辑共享文件。
成员（Member）：使用被分配的 Agent。	查看者（Viewer）：仅可对话并阅读共享文件。

Claude Code 或 Codex：使用您自己的订阅

Claude Code 运行在您的 Anthropic 订阅上，Codex 运行在您的 ChatGPT 订阅上。每位用户均可接入自己的订阅，同时也支持运行在自有硬件上的本地模型。您可以针对每个 Agent 选择底层引擎，甚至可以在单次对话中灵活切换。

在你离开时，它们依然在工作

Agent 可以按任意定时计划运行、在 Webhook 事件触发时启动，或在您指定的单次时间点执行。它们会在您的工作区内开展工作，输出报告、文件与更新，并在需要人工介入时向您发送通知。通知分为四级严重程度，从轻微的提示音到持续的危险警报不等。每次运行都是一次完整的对话，您可以随时打开、查阅并继续推进，无需人工实时盯盘。

为 Agent 分配一个电话号码

Agent 能够接听和拨打电话。支持接入您自有的 Twilio 账号，或连接您现有的 Asterisk、FreePBX 服务器。

运行在您的服务器上，作业于您的设备之上

默认情况下，所有 Agent 均在您服务器的隔离沙箱中运行，并由统一的控制台进行调度控制。您只需一行安装命令，即可将运行 macOS、Linux 或 Windows 的笔记本、工作站或 PC 与服务器配对。配对设备仅与您的服务器保持单向出站连接，因此无需开放任何公网端口，文件也能保持实时同步。同一个 Agent 随后可以直接在该设备上运行并获得完整系统访问权限，您可以在任意地点通过同一个控制台进行操控。若该设备离线，服务器将自动接管任务。

属于您自己的 Agent 云端架构

视频/嵌入内容：hero-v8-web.mp4

整段视频完全由一个 OtoDock Agent 负责导演、录制与剪辑。在 otodock.io 查看完整画质。

部门机制（Departments）。 将您的 Agent 划分为不同部门，设定相互之间的委托权限，或者组织它们召开联合会议。	实时仪表板（Live dashboards）。 每个 Agent 都配有实时仪表板，助您轻松掌控全局。
海量工具（Endless tools）。 您的 Agent 就是数字化员工。它们能接入海量工具，从日历到智能家居无所不包。	文档处理（Documents）。 直接在对话界面中原生编辑 Excel、Word 和 PowerPoint 文件。
Artifacts 交互组件。 Agent 能以实时交互式 UI 作出响应，直接渲染在对话窗口中。	协同会议（Meetings）。 专业领域 Agent 共同探讨某个议题并在控制台中实时协作推导出解决方案。
终端（Terminal）。 Claude Code 或 Codex 终端直接在控制台中唤起，并已预载该 Agent 的所有工具。	远程机器（Remote machines）。 同一个 Agent 及其所有工具和同步工作区，可通过单条命令在任何配对机器上运行，并调用该机器上的全部系统能力。

Agent 功能强大，因此 OtoDock 在设计上默认假定它们不可盲目信任。每个服务端 Agent 均运行在具备持久网络隔离的内核沙箱内部，且需要管理员按服务逐步授权。

• 围绕每个 Agent 的内核级沙箱。 每个会话运行在专属的挂载和进程命名空间中。目录会根据用户在每个 Agent 中的对应角色自动挂载。
• 网络隔离。 私有 IP 段、内网 LAN 以及云厂商元数据端点在设计上默认不可达。需要访问本地服务的 MCP 工具可由管理员按 Agent 精准分配权限范围。
• 安全凭证管理。 凭据在静态存储时处于加密状态，且仅在单个会话中按需注入。Agent 可以使用凭据，但永远无法窥见明文内容。
• 团队就绪。 从首次安装起即标配 SSO 单点登录、双因素身份验证（2FA）以及单用户费用预算限制。

阅读安全模型规范 →

单一平台，全套工具箱：

• 持久化记忆： Agent 为每位用户维护透明、可编辑的记忆文件。
• 文档与文件处理： Agent 支持创建和编辑 Word、Excel、PowerPoint 及 PDF 文件，并且每个文件都能直接在聊天界面的实时编辑器中打开。
• 图像处理： Agent 能够在对话中生成和编辑图片，并内置专业级管线处理照片编辑。
• 视频与音频： Agent 可生成素材和转场、剪辑时间线、添加字幕、旁白与背景音乐，将音视频转录为文本与字幕，并自主生成语音与音乐。
• 网页浏览： 来自社区目录的浏览器工具允许 Agent 对实时网络进行检索调研。
• 内置工具集： 平台自带定时调度、触发器、通知系统、协同会议、任务委派、电话呼叫、跨 Agent 文件传输、实时图表与迷你应用、SSH 主机管理，以及对配对机器的浏览器与系统级控制。
• 社区生态目录： 开箱即用的 Agent、工具与技能支持一键安装，覆盖 GitHub、Notion、Home Assistant、Nextcloud、Prometheus、UniFi 和 Uptime Kuma 等，并持续更新。
• 原生可扩展性： 支持根据 Manifest 安装任意 MCP（Model Context Protocol）工具服务器，并按 Agent 进行颗粒化指派。
• 用量与预算管控： 成本按用户和 Agent 双重维度追踪，支持配置周限额或月度预算。
• 企业级安全基线： 首发即支持 SSO 与 OIDC、双因素身份验证、分 Agent 角色权限体系、加密凭据以及限定权限范围的 API Key。

查看完整功能特性请访问 otodock.io/features。

仅需一台安装了 Docker 的 Linux 服务器即可开始部署（最低内存要求 4 GB，详见硬件规格指南）。

OtoDock — Collaborative Agents

The brains of your company, built on Claude Code & Codex, working on your Anthropic and OpenAI subscriptions.

Docs · Install · Features · Discussions

Self-hosted · Multi-tenant by design · Fair source
Runs on Claude Code · Codex · your API keys · local models

Dashboard highlights.The two-minute video tells the whole story.

If OtoDock is useful to you, a star on this repository helps other people find it.

---

OtoDock acts as the brain of your company. You create powerful agents that connect to the tools your company runs on, work in departments, delegate to each other, and keep working on their own when no one is watching. It is multi-tenant by design. Many people work with the same agents, and there are four modes that decide how an agent is shared and where its work lands. You work with your agents on your self-hosted dashboard, or you can even give them a phone line and talk to them. All of it runs on your own Anthropic and OpenAI subscriptions, and on local models.

Every agent gets a name and a job, like a Personal Assistant, a System Admin or a Marketing Manager, and runs on your server. An agent is made of six parts, and you can edit every one of them.

• Persona. Plain-language instructions tell the agent who it is and how it works.
• Memory. The agent keeps its own notes across chats.
• Workspace. The folder where the everyday work lands. It is private for each person, shared by the whole team, or both, depending on the agent's mode.
• Knowledge. Reference documents are always on hand.
• Skills. The agent learns techniques it can apply to its work.
• Tools. The agent uses only what it is allowed to use.

Four ways to share one agent

Every agent has a mode that decides how people share it.

Mode	What it means
Personal only	A private workspace for each person.
Personal + shared	A private workspace for each person as the default, and a shared one for the team.
Shared + personal	A shared team workspace as the default for everyday work, plus a private one per person.
Shared only	Everyone shares one history and one workspace.

Everyone gets the right seat

Admins add each person to the agents they need.

Three platform roles	Three roles on every agent
An Admin runs the platform.	A Manager has full control of the agent.
A Creator creates agents.	An Editor edits the shared files.
A Member uses the agents they are given.	A Viewer chats and reads the shared files.

Claude Code or Codex. Your subscription.

Claude Code runs on your Anthropic subscription, and Codex runs on your ChatGPT subscription. Every user connects their own subscription. Local models on your own hardware work as well. You pick the engine per agent, and you can switch it per chat.

They work while you are away

Agents run on any schedule, when a webhook event fires, or once at a time you choose. They work in your workspace, where their reports, files and updates land, and they notify you when something needs you. Notifications come in four severities, from a quiet chime to a persistent danger alarm. Every run is a full chat you can open, read and continue, so no one has to be watching.

Give an agent a phone number

Agents answer and place phone calls. Bring your own Twilio account, or connect the Asterisk or FreePBX server you already run.

Runs on your server. Works on your machines.

By default, all your agents run on your server, in an isolated sandbox, with one dashboard controlling all of them. You can also pair a laptop, a workstation or a PC running macOS, Linux or Windows with a one-line install. The machine keeps a single outbound connection to your server, so it needs no open ports, and its files stay in sync. The same agent then runs on that machine with full access to it, and you work with it from the same dashboard, anywhere. If the machine goes offline, your server takes over.

Your own cloud of agents.

hero-v8-web.mp4

This entire video was directed, captured and edited by an OtoDock agent.Watch it in full quality on otodock.io.

Departments. Organize your agents into departments, decide who can delegate to whom, or put them in a meeting together.	Live dashboards. Every agent gets live dashboards, so you can manage them effortlessly.
Endless tools. Your agents are digital employees. They connect to endless tools, from your calendar to your smart home.	Documents. They edit Excel, Word and PowerPoint files right in the chat.
Artifacts. Agents answer with live, interactive UI, rendered right in the chat.	Meetings. Specialist agents discuss a topic together and converge on an answer, live in the dashboard.
Terminal. The Claude Code or Codex terminal opens in the dashboard, with the agent's tools already loaded.	Remote machines. The same agent, with all its tools and its synced workspace, runs on any machine you pair with one command, and it can use everything on that machine.

Agents are powerful, so OtoDock assumes they can't be trusted. Every server-side agent runs inside a kernel sandbox with always-on network isolation, and you grant access one service at a time.

• A kernel sandbox around every agent. Each session runs in its own mount and process namespace. Folders are mounted automatically from each user's role per agent.
• Network isolation. Private ranges, your LAN, and cloud metadata endpoints are unreachable by design. MCP tools that need a local service can be granted scoped access by the admin, per agent.
• Secure credentials. Credentials are encrypted at rest and injected only per session. Agents can use them, but never see them.
• Ready for teams. SSO sign-in, two-factor auth, and per-user cost budgets come standard, from the first install.

Read the security model →

One platform, the whole toolkit.

• Memory that persists. Agents keep transparent, editable memory files, per user and per agent.
• Documents and files. Agents create and edit Word, Excel, PowerPoint and PDF files, and every file opens in a live editor inside the chat.
• Images. Agents generate and edit images in the chat, and a professional-grade pipeline handles photo editing.
• Video and audio. Agents generate footage and transitions, cut timelines, add captions, voice-overs and music, transcribe audio and video into text and subtitles, and produce speech and music of their own.
• Web browsing. The browser tool from the community catalog lets agents research the live web.
• Built-in tools. Schedules, triggers, notifications, meetings, delegation, phone calls, file transfer between agents, live charts and mini-apps, SSH hosts, and browser and computer control on paired machines all ship with the platform.
• Community catalog. Ready-made agents, tools and skills install in one click, with GitHub, Notion, Home Assistant, Nextcloud, Prometheus, UniFi and Uptime Kuma among them, and more landing regularly.
• Extensible by design. Any MCP tool server installs from a manifest, and tools are assigned per agent.
• Usage and budgets. Costs are tracked per user and per agent, with weekly or monthly limits.
• Team-ready security. SSO and OIDC, two-factor auth, per-agent roles, encrypted credentials and scoped API keys come from day one.

See all features at otodock.io/features.

A Linux server with Docker is all you need (4 GB RAM minimum, see the sizing guide). Create a folder for the install, then run the script in it:

mkdir otodock && cd otodock
curl -fsSLO https://raw.githubusercontent.com/OtoDock/oto-dock/main/scripts/install.sh
bash install.sh

The installer checks Docker, writes a .env with a generated database password, handles the Ubuntu 24.04+ host step automatically when the host needs it, downloads the release-pinned docker-compose.yml plus the phone-service overlay, and starts the stack. Everything lands in the folder you run it from. The script performs fresh installs only, and stops rather than touch an existing install.

If your users browse to the server by name or IP, set DASHBOARD_PUBLIC_URL in the generated .env. Behind a reverse proxy, also set TRUSTED_PROXY to your proxy's IP (reverse proxy & HTTPS). Every optional knob is documented in the Configuration reference, and the installation guide also covers building from source, bare-metal development, and running behind a reverse proxy with HTTPS.

Open http://localhost:8400 and the setup wizard greets you. Create the owner account, and OtoDock installs the Personal Assistant for you, with the tools it needs. Connect your Claude or ChatGPT subscription under Setup, AI Engines, and the banner that reminds you goes away. Open the Personal Assistant from the Agents page and send it your first message. Its reply streams in live, with every tool call and every file it creates (First run). From there you add agents from the community catalog or build your own, and you add the people who will work with them.

 dashboard/   React dashboard — chat, agents, tasks, files, admin
 proxy/       Platform core (FastAPI) — sessions, security, scheduling,
              the agent sandbox, and the WebSocket hub the dashboard talks to
 mcps/        MCP tool servers: OtoDock's custom set (files, memory, tasks,
              meetings, notifications, …) + community mirrors
 audio/       Speech package — STT / TTS / voice activity, provider-agnostic
 phone/       Telephony daemon — live calls over Twilio or FreePBX/Asterisk
 satellite/   Remote-machine agent — pairs your own hardware to the platform
 scripts/     Install, compose, backup/restore, and maintainer tooling

Agents run as Claude Code / Codex processes inside per-session kernel sandboxes, talk to their tools over MCP, and stream every step back to the dashboard. PostgreSQL holds the platform state.

Community

Contributions are welcome: see CONTRIBUTING.md. Security reports: SECURITY.md.

OtoDock is fair source : licensed under the Functional Source License, v1.1, with Apache 2.0 future grant (FSL-1.1-Apache-2.0). You can use, run, modify, and redistribute it for anything except competing with OtoDock commercially — and each version automatically becomes plain Apache 2.0 two years after its release.

Self-hosting is free for up to 5 users. Growing teams license by seats. Same software, signed key, no feature gates on your data or your agents.

OtoDock started with Claude Code. I wanted to use it beyond the terminal and connect it to my services. So I built the platform around that idea. It has since become the place where my own agents work: large parts of OtoDock were built, tested, and shipped by agents running on OtoDock.

It's fair source so you can run it the same way for your company, and you can share your builds in the discussions.

— Dimitris Mourtzis