Ownward

Your work, carried forward.

English · 简体中文

AI sessions end. Development work should not have to start over.

Ownward is a local AI workbench for developers. Dispatch work from your phone, let Claude Code, Codex, or CodeBuddy take turns, and bring the execution facts from every agent back into one project memory.

Ownward is early alpha software. The command-line entry point is own.

Once AI is part of your development workflow, the hard problem is usually not that a model cannot write code. The hard problem is continuity: you leave your desk and lose the thread; Claude hits a limit and switching to Codex means explaining everything again; several CLIs change the same project with no single record of what happened; and writing a status update means searching through chat logs and Git history all over again.

Ownward runs on your Mac and keeps tasks, project directories, and work records together. Android and iPhone act as remote workbenches. Your code, credentials, and agent CLIs stay on the Mac.

Five ways Ownward helps with development

1. Code from Android or iPhone without adding another provider

Connect your phone to Ownward on the Mac to dispatch a development task, choose the project directory, add extra directories, select a model, and set access permissions. You can watch streaming replies, tool calls, and images; ask a follow-up question; interrupt a task; switch engines; and inspect repository status, diffs, tests, commits, and pull requests.

The Android client is available from GitHub Releases. The iPhone client requires iOS 26+ and can be installed on a device with Xcode. TestFlight distribution is provided by the maintainer. See ios/README.md for build instructions.

Neither mobile client copies your code to the phone for execution. Remote access uses a token. Put Ownward behind Tailscale or a trusted TLS proxy instead of exposing the local port directly to the public internet.

Ownward can use Claude Code, Codex, and Tencent CodeBuddy. Configuring only CodeBuddy still gives you the complete task workflow without access to an overseas provider. Ownward is not a proxy and does not change the network requirements of Claude or Codex.

2. Continue task sessions and ordinary chats across engines

When you switch the engine for a task session, Ownward creates a traceable successor session. The new engine receives a bounded slice of recent history and inherits the original working directory, extra directories, and access level. The handoff prompt tells it to inspect the actual Git and file state instead of replaying old tool calls. The previous chain stays available for review, and an operation with an unknown result must be confirmed by a person before another engine takes over.

Claude Code --rate limit--> Codex --different capability--> CodeBuddy
      `----------- the old session and execution record stay available --------'

Ordinary Chat switches differently. Change the provider in an existing conversation and Ownward replays the chat history as context. It does not create a task successor chain and does not carry task workspace state. That is useful for discussion and advice. Use a task session when the conversation needs to operate on real code.

3. Put all three agents' execution facts back in one project

Tasks dispatched to Claude Code, Codex, or CodeBuddy create the same Run and Flight Record structure. The goal, execution trail, code changes, commit, and conclusion return to the project in one format, no matter which agent handled the turn.

4. Harvest sessions into durable project memory

Substantive sessions started in external Claude Code or Codex CLI are discovered and harvested automatically. CodeBuddy's private transcript cannot currently be read from outside, so only CodeBuddy tasks started by Ownward are harvested.

Harvested material is written to a local Markdown vault: recent material goes to the inbox, project evolution goes to the log, and task facts go to flights. When the model suggests long-term knowledge, it first writes to _candidates/; a person confirms it before it becomes official project memory. The next task can reuse those records instead of hoping an agent remembers the last conversation.

5. Draft recurring updates from the work that already happened

Routines fit recurring responsibilities such as stand-ups, weekly reports, and project syncs. At the scheduled time, Ownward gathers recent harvested records, task results, and project memory into a draft for review.

A person must approve the draft before it is written to a Feishu document. Ownward gathers and drafts the material, but it does not skip the final publishing decision. Routines are off by default. Start with examples/routines.json when configuring data/routines.json; see the configuration guide for Feishu setup and the rest of the configuration surface.

Quick start

The server currently runs on macOS and needs:

• Bun
• Git
• A logged-in Claude Code or Codex CLI
• Optional: Tencent CodeBuddy

git clone https://github.com/whtis/ownward.git
cd ownward
./install.sh
open http://127.0.0.1:4517

The installer creates local configuration and starts the daemon and an isolated Runner through launchd. The workbench listens on 127.0.0.1:4517 by default.

After the first launch, open Settings → AI Engines to select the background decision engine, enable or disable Claude Code, Codex, and CodeBuddy, and choose the default project directory, provider, model, and permission for new tasks. For a Codex-only installation, select Codex there. CodeBuddy is off by default and can be enabled from the same page.

How work gets carried forward

Start a task from phone, Web, or Terminal
                    |
         Claude / Codex / CodeBuddy
            hand off when needed
                    |
          Run / Flight Record / diff
                    |
          Harvest into project material
                    |
       A person promotes candidate knowledge
              into the current truth
                    |
        The next task and Routine reuse it

This loop is not about saving more chat. It gives the next development session the facts left by the last one. The vault is ordinary Markdown that you can search, edit, commit to Git, or manage with Obsidian. The default location is ~/Documents/ownward-vault/.

The same Kernel can host other workbenches

The development workbench is Ownward's built-in dev Vertical. Its Kernel, independent Runner, provider adapters, fact storage, and permission layer do not depend on the specific domain of writing code.

The Vertical contract defines controlled capabilities such as sessions, tasks, actions, storage, scheduler, LLM, and sources. A built-in Vertical can request capabilities through that contract. The external Host currently exposes storage, actions, scheduler, LLM, and sources; sessions and tasks remain blocked by the Host availability gate. An external Vertical can also mount its own API routes, pages, and navigation, and can run in a separate Host process for crash isolation and development-time hot reload.

The included Desk extension is aimed at recruiting workflows. It uses the same engine chain, Actions, scheduler, storage, and sources while keeping the full domain implementation outside this repository. The public read-only Vertical example uses scoped storage to show a sanitized candidate list and verify Host, routing, page, and authorization boundaries.

The same approach works for other personal workbenches. An external Vertical must be trusted local code explicitly enabled by the user. A separate process isolates crashes and lifecycle, but it is not a security sandbox. See the development guide for the extension contract and development rules.

Capabilities and boundaries

Capability	Claude Code	Codex	CodeBuddy
Dispatch, continue, and supervise inside Ownward	✓	✓	✓
Cross-engine task-session handoff	✓	✓	✓
Continue ordinary Chat across engines	✓	✓	Configurable
Put Ownward tasks into the unified record	✓	✓	✓
Harvest sessions from external CLIs	✓	✓	—
Work without an overseas provider	—	—	✓

Ownward is not an IDE or an agent security sandbox. It keeps work continuous; it does not replace Claude Code, Codex, CodeBuddy, or Git. It also cannot yet import the complete history from ChatGPT or the Claude web app.

Dispatch a task

Use the workbench or the CLI:

# Start Claude Code in Terminal so you can take over at any time
bin/own work ~/workspace/example "Fix the login-page crash"

# Run in an isolated worktree in the background
bin/own work ~/workspace/example "Add missing unit tests" --bg
bin/own work ~/workspace/example "Refactor the utils directory" --bg --codex

bin/own tasks
bin/own done <id>

Background tasks run in an isolated worktree by default. A worktree reduces the risk of changing the main checkout by accident, but it is not a permission sandbox.

Workbench tabs

• Today: Actions, agent wrap-up, Routines, and calendar items that need attention
• Tasks: Dispatch, observe, ask follow-up questions, approve, take over, or hand off a task
• Chat: Ordinary Chat, with provider switching inside an existing conversation
• Feed: Triage results from external events
• Notes: Browse and edit the Markdown vault
• System: Event sources, schedules, logs, and runtime status
• Settings: Configure providers, event sources, notifications, automation, remote listening, and task-dispatch defaults with a reviewable diff

GitHub, Feishu, Gmail, and stock event sources are off by default. Without any external account, you can still use task dispatch, session harvesting, project records, Actions, Heartbeat, notes, and macOS notifications.

Configuration and security

Use the workbench's Settings tab for everyday configuration. It covers owner identity, timezone, vault, providers, event sources, Heartbeat, digest, notifications, Dashboard listening, and the default directory, provider, model, and permission for new tasks. Ownward shows a deterministic diff before approval, then writes the override and restarts the Runner and daemon through the paired release transaction.

Underneath, Ownward still uses two configuration layers. The checked-in config.default.json contains defaults, while the local config.json contains only overrides and is never committed. Prompt editing and new Routines still use:

• prompts/owner.md: which people and events matter, plus writing preferences;
• prompts/heartbeat.md: the proactive check list;
• data/routines.json: recurring responsibilities.

If you edit config.json directly, run bash install.sh so the Runner and daemon switch to the same frozen configuration snapshot. Applying a change from Settings runs that transaction automatically. See the full configuration guide.

Local storage and a local control plane do not mean the model runs locally. Content sent to a provider is covered by that provider's terms, and an agent may run commands or change files with the current user's permissions. Never commit config.json, prompts/owner.md, data/, the vault, credentials, or raw transcripts to Git.

The Dashboard validates Host and Origin headers. See SECURITY.md for the security boundary and reporting process.

Architecture and development

Ownward is a Bun / TypeScript daemon with an embedded static Web workbench and no runtime npm dependencies. Provider tasks run in an independent Runner. During a daemon update or restart, the Runner drains safely instead of marking active tasks as successful.

Read the architecture guide and the development guide before changing the project. Start development with:

bun install --frozen-lockfile
./verify.sh

The verification gate builds the project, runs the TypeScript type checker and unit tests, smoke-tests an isolated daemon, probes the API, and parses the Web JavaScript. See CONTRIBUTING.md for the contribution process.

Ownward is licensed under the Apache License 2.0.

Remote access from a phone

Open Settings → Advanced → Dashboard, change Listening scope to LAN access, review the high-risk diff, and apply it. If you only have terminal access, set the same option manually:

{
  "dashboard": { "listen": "all" }
}

bash install.sh

Ownward will listen on 0.0.0.0:4517 and require a remote access token. Use a firewall or router to block direct public access to port 4517, and expose only the HTTPS entry point below. The first phone visit asks for the token stored in data/secrets/api-token.txt.

Nginx

Configure a valid TLS certificate for your domain, then add a server block like this:

server {
    listen 443 ssl;
    server_name ownward.example.com;

    # The first login uses a query parameter. Do not write the full request to access logs.
    access_log off;

    ssl_certificate     /path/to/fullchain.pem;
    ssl_certificate_key /path/to/privkey.pem;

    location / {
        proxy_pass http://127.0.0.1:4517;
        proxy_http_version 1.1;
        proxy_set_header Host $host;
        proxy_set_header X-Forwarded-For $proxy_add_x_forwarded_for;
        proxy_set_header X-Forwarded-Proto https;
        proxy_buffering off;
        proxy_read_timeout 3600s;
    }
}

Reload Nginx, then open https://ownward.example.com on the phone. X-Forwarded-For makes Ownward apply token authentication to the remote request. X-Forwarded-Proto: https makes the login cookie use Secure.

Cloudflare Tunnel

Cloudflare Tunnel does not require an inbound router port. Install cloudflared and prepare a domain connected to Cloudflare:

cloudflared tunnel login
cloudflared tunnel create ownward
cloudflared tunnel route dns ownward ownward.example.com

Put the Tunnel UUID and credentials path returned by those commands in ~/.cloudflared/config.yml:

tunnel: <TUNNEL-UUID>
credentials-file: /Users/you/.cloudflared/<TUNNEL-UUID>.json
ingress:
  - hostname: ownward.example.com
    service: http://127.0.0.1:4517
  - service: http_status:404

Start and verify the tunnel:

cloudflared tunnel run ownward

After confirming that the phone can connect, follow the Cloudflare documentation to register the Tunnel as a persistent macOS service.

The first login briefly uses a token in a query parameter, then exchanges it for an HttpOnly cookie. Do not log query strings, copy this login URL, or capture it in a screenshot. Exclude query parameters from Cloudflare logs, monitoring, and analytics too. If the token may have leaked, delete data/secrets/api-token.txt and run bash install.sh again so every client must log in again.

If you mainly use a browser, you can also enable Cloudflare Access. Before connecting the native Android or iOS clients, confirm that their authentication flow is compatible with the Access policy. Keep Ownward's own access token even when Access is enabled. See the Nginx reverse-proxy documentation for the proxy directives.

Ownward（简体中文）

Ownward 是给研发者使用的本地 AI 工作台：从手机派发任务，让 Claude Code、Codex 或 CodeBuddy 接力工作，再把每个 Agent 的执行事实收回同一个项目记忆。

项目目前仍处于 early alpha，命令行入口是 own。Ownward 常驻在你的 Mac 上，统一 管理任务、项目目录和工作记录；Android / iPhone 作为远程工作台，代码、凭据和 Agent CLI 仍留在 Mac 上。

主要能力

1. 手机派发和旁观研发任务：选择项目目录、附加目录、模型和权限，查看流式回复、 工具调用、图片、diff、测试、commit 和 PR，也可以继续追问、中断或接管。
2. 跨引擎接力：任务会话切换引擎时建立可追溯的后继会话，带上有界的近期历史、 工作目录和权限；新引擎必须先检查真实 Git / 文件状态，不能重放旧工具调用。
3. 统一执行记录：Claude Code、Codex 和 CodeBuddy 的任务都写入同一套 Run 和 Flight Record，目标、过程、代码变化、commit 和结论不会散落在不同工具里。
4. 会话收割和项目记忆：外部 Claude Code、Codex CLI 的实质会话会自动收割到 Markdown vault；长期知识先进入 _candidates/，经人确认后才成为正式记忆。
5. Routine 起草：从近期记录、任务结果和项目记忆生成晨会、周报或项目同步草稿； 写入飞书文档前必须由人确认。

快速开始

服务端目前支持 macOS，需要 Bun、Git，以及已登录的 Claude Code 或 Codex CLI。 CodeBuddy 可选。

git clone https://github.com/whtis/ownward.git
cd ownward
./install.sh
open http://127.0.0.1:4517

安装脚本会生成本机配置，并通过 launchd 启动 daemon 和独立 Runner。默认只监听 127.0.0.1:4517。首次打开后进入「设置 → AI 引擎」，可以切换后台决策引擎、启停 Claude Code / Codex / CodeBuddy，并配置派发任务时默认使用的目录、引擎、模型和权限。

Vertical 扩展

研发工作台是内置的 dev Vertical。Kernel、Runner、Provider 适配器、事实层和权限层 不依赖“写代码”这一具体领域。外部 Vertical 可以通过能力契约挂载自己的 API、页面和 导航，并在独立 Host 进程中运行。公开的 只读 Vertical 示例 用 scoped storage 展示脱敏 候选人列表，验证 Host、路由、页面和授权边界。

外部 Vertical 必须是用户明确启用的 trusted 本地代码。独立进程用于隔离崩溃和生命周期， 不是安全沙箱。扩展契约和开发规则见开发指南。

配置与安全

日常配置优先在工作台「设置」完成。Owner、时区、Vault、Provider、事件源、心跳、日报、 通知、Dashboard 监听范围和任务派发默认值都可以在页面修改；应用前会展示 diff 并要求确认。 底层仍由 config.default.json 和本机 config.json 两层组成。提示词和新增 Routine 继续通过 prompts/owner.md、prompts/heartbeat.md 与 data/routines.json 管理。直接手改配置后需要 运行 bash install.sh，设置页应用则会自动执行同一套事务。

Ownward 不是 IDE，也不是 Agent 安全沙箱。Agent 可能以当前用户权限执行命令和修改文件， 发给 Provider 的内容受对应服务条款约束。不要提交配置、凭据、vault、原始 transcript 或 运行时数据。远程访问请放在 Tailscale 或可信 TLS 代理之后，并保留 Ownward 自己的令牌鉴权。 完整配置见配置指南，安全边界见 SECURITY.md。

开发与验证

bun install --frozen-lockfile
./verify.sh

验证门包含构建、TypeScript 类型检查、单元测试、daemon 冒烟、API 探活和 Web JavaScript 解析。贡献流程见 CONTRIBUTING.md，架构说明见 docs/architecture-v1.md。

手机从外网连接

在「设置 → 高级 → Dashboard」把监听范围改为“局域网可访问”并应用，再用 Nginx 或 Cloudflare Tunnel 提供 HTTPS 入口。没有页面时也可以手动把 dashboard.listen 设为 "all" 后运行 bash install.sh。不要直接把 4517 端口暴露到公网；首次登录使用 data/secrets/api-token.txt 中的令牌，换取 HttpOnly cookie 后即可使用。

---

Ownward - your work, carried forward.