项目的 README,来自 github.com/mdiri-ai/SoduWork。合并为一个 markdown 文件:README.md。
English · 简体中文
A chat ends with an answer. A SoduWork request ends when the person who asked signs off.
Multi-harness · Multi-agent · Self-hosted
Works with Pi · OpenCode · DeepSeek Harness · Qwen Code · Kimi Code · Codex Harness · Codex
git clone https://github.com/mdiri-ai/SoduWork
cd SoduWork
| Platform | Start command |
|---|---|
| macOS / Linux | sh start.sh --desktop |
| Windows PowerShell | .\start.cmd --desktop |
| Windows Command Prompt | start.cmd --desktop |
Website | Live demo | Documentation | Docs website | Releases | Discord
| One harness does the work, another reviews it | Pick Pi to do the work. OpenCode, on a different model, reviews it against the acceptance criteria. If the review fails, the work goes back with the reasons. |
| Hand a request to a colleague's agent | People and AI work together: you file a request, AI writes the acceptance criteria, and you assign it to yourself or a colleague. The assignee's agent does the work and reviews it, and a person signs off. Everyone on the request can see who did what. |
| Turn one person's know-how into a team asset | When you and AI get a task right, save the chat as a plugin. After review, anyone on the team can use it and contribute improvements, with access granted as needed. |
|
🔀 Smart dispatch The right agent and model for each job.
|
🛡️ Built-in governance Your organization decides which AI does what, and where.
|
✅ Verified delivery Set the bar first. Done means the requester signed off.
|
SoduWork is in public beta. Try it and tell us what you think, and evaluate it before using it in production.
You need Python 3.12+, Node.js 22.19+ and Git.
From the cloned repository, run sh start.sh --desktop on macOS/Linux, .\start.cmd --desktop in Windows PowerShell, or start.cmd --desktop in Command Prompt.
The first run installs dependencies, then starts the Hub, the admin console and the desktop client on this computer. Set a password for the admin account (12–128 characters) when asked. Then add a model provider under Model catalog in the admin console, start a chat in the client and try a small task. Pi is built in.
--desktop when starting from source. The browser admin console configures the service; check that Connection on the desktop sign-in screen points to that same Hub.For the first message, use an OpenAI-compatible chat model with Pi. A successful connection probe does not guarantee a successful inference: balance, model permissions and protocol support still matter. If the Agent list is empty, check Agent management and the Runner connection. If the model list is empty, check that the model exists, is enabled, and both interfaces use the same Hub.
See Run from source: send your first AI message for field-by-field setup and troubleshooting.
Download the client from Releases, enter https://service.soduwork.ai under Connection on the sign-in screen, then sign up and sign in.
The demo is for evaluation only. Don't upload real business data or anything sensitive.
Follow the illustrated desktop installation and first conversation guide.
| OS | Package |
|---|---|
| Windows | SoduWork-<version>-Setup.exe |
| macOS (Apple Silicon / Intel) | SoduWork-<version>-mac-<arch>.dmg |
| Linux | SoduWork-<version>-linux.deb / .rpm / .tar.gz |
Download from Releases. On first launch, enter your Hub address under Connection, then sign in. On macOS, if you see "cannot verify the developer", open System Settings → Privacy & Security and choose Open Anyway.
With Docker installed and running, from the repository root:
python3 ops/self-host/deploy.py up
python3 ops/self-host/deploy.py credentials # show the initial admin password
This starts the Hub, the admin console and PostgreSQL on one server. Put an HTTPS reverse proxy in front of it to open it to your team. See Deploy with Docker and Deploy for a team.
Ports, data folders and troubleshooting are in Run from source and the FAQ.
AI agents give everyone fast output. What no one gets for free is trust in that output — or a way to keep what worked. That's where individuals and teams get stuck alike:
| The problem | What SoduWork does |
|---|---|
| Fast output, slow trust — whether you work alone or in a team, the time AI saves goes back into checking what it produced | The request is the spec. Criteria set before work starts; a separate AI review checks each one; a person signs off |
| What worked dies in a chat thread — for one person, it's last month's perfect run you can't reproduce; for a team, it's one person's method nobody else can use | Get it right once, keep it as a plugin. Reuse it yourself or publish it for everyone — every improvement reaches everyone |
| A drawer full of AI tools — three subscriptions and no standard for which one does what | One workbench, your playbook. Your plugins run on every harness. Switch mid-conversation, keep your history and files |
| Results that never enter the workflow — output gets copy-pasted around, and every handoff re-explains the background | Issues that carry the context. People and agents pick up where it left off, with a full record of who did what |
Requests, execution, review and reuse all happen in one shared issue pool:
flowchart LR
A[Raise a request] --> B[AI refines it<br/>adds acceptance criteria]
B --> C[Assign<br/>yourself · a colleague · an agent]
C --> D[Agent runs it<br/>with a team plugin]
D --> E[AI review<br/>assignee confirms]
E --> F[Requester signs off]
F --> G[Save as a plugin]
G -.reused by the team.-> D
| Role | Where they join |
|---|---|
| Solo developer | Puts requests in the pool, lets agents work through them, and only signs off |
| Specialist | Gets a task right with AI, saves the method as a plugin, and publishes it to the team |
| Manager | Drops in a request; AI fills in the criteria and routes it to a person or an agent; signs off at the end |
| New team member | Uses the plugins the organization has built up to do the work to team standards, and improves them when they find a gap |
Who asked, who did the work, which plugin ran and who approved it is always on record.
A harness is the agent runtime that does the work. It is not the model. SoduWork connects every harness through one adapter interface, so issues, plugins and review don't depend on any single harness. Switch harnesses mid-conversation and keep the history and the working folder. Programs that support the Agent Client Protocol (ACP) can be added with a catalog declaration.
| Harness | Models | How it connects |
|---|---|---|
| Pi | Your SoduWork models | Built in |
| OpenCode | Your SoduWork models | Native adapter, installed on demand |
| DeepSeek Harness | Your SoduWork models, or your own DeepSeek key | ACP, installed on demand |
| Qwen Code | Your SoduWork models | ACP, installed on demand |
| Kimi Code | Your SoduWork models | ACP, installed on demand |
| Codex Harness | Your SoduWork models | ACP, installed on demand |
| Codex | Its own OpenAI account | Native adapter, installed on demand |
Agents work together at several levels:
| Level | Who decides | Example |
|---|---|---|
| Harness | The harness itself | OpenCode or Qwen Code starts its own subagents to explore the code or split a task |
| Single method | A Skill or plugin sets the method, process, rules and what counts as done | A review plugin tells the agent which checks to run and what evidence to attach |
| Combined methods | A plugin workflow picks the harness, Skill and model for each step | Pi drafts, Qwen Code checks the sources, and the run stops if the check fails |
| People and agents | The requester and the assignee or their agent | A request goes to a colleague, their agent does the work, the requester signs off |
| In parallel | The Hub | Different issues run on different agents at the same time, including on cloud runners |
Between steps and between people, work moves the way the plugin and the assignee say. Agents don't hand work to other agents on their own.
A skill tells AI how to do something. An agent wraps AI into an assistant. SoduWork plugins and apps turn the way your team works into something that runs, gets verified, improves and stays governed.
| Skill | Agent | SoduWork plugin | SoduWork app | |
|---|---|---|---|---|
| What it is | A method doc that AI reads and follows | Prompts, knowledge, tools or a platform flow | A versioned package: method, inputs, acceptance criteria, workflow | A plugin with its own UI and an AI chat panel |
| Best for | A single technique | One type of task | Standard team work, run by AI | Repeated work in a dedicated UI |
| How you use it | AI picks it up | Inside its platform | / in chat · on issues · in automations |
Its own window + AI sidebar |
| Step-by-step process | ✗ | ◐ | ✓ | ✓ |
| Choose and switch harness | ✗ | ◐ | ✓ | ✓ |
| Criteria, AI review, sign-off | ✗ | ◐ | ✓ | ✓ |
| Versioning | ✗ | ◐ | ✓ | ✓ |
| UI and interaction | ✗ | ◐ Chat | ◐ Chat | ✓ |
| Review, grant, revoke | ◐ Configurable | ◐ Configurable | ✓ | ✓ |
✓ Supported ◐ Partial ✗ Not supported. "Agent" means an assistant built on a typical agent platform; capabilities vary by platform.
From your chat to your playbook — yours or the team's: create or save a plugin in chat → pre-check and try it locally → keep it for yourself, or submit it for review and publish it to the team catalog. Solo use stays light; team sharing adds review, grants and versioning.
{
"schemaVersion": 1,
"key": "weekly-brief",
"version": "1.0.0",
"name": "Weekly brief",
"workflow": {
"nodes": [
{ "key": "draft", "type": "ai", "config": { "agentKind": "local:pi", "prompt": "Write deliverables/brief.md from this week's tasks." } },
{ "key": "check", "type": "ai_review", "dependsOn": ["draft"],
"config": { "agentKind": "local:qwen", "prompt": "Check each claim against the sources. Block if any claim has no source." } },
{ "key": "done", "type": "publish", "dependsOn": ["check"] }
]
}
}
You don't have to write this by hand: describe the work to Plugin Developer in a chat and it produces the file. See the plugin guide and apps.
| Low-code agent builders | Coding agent harnesses | Multi-agent orchestrators | AI office assistants | SoduWork | |
|---|---|---|---|---|---|
| Domain experts build and refine agents in chat | ◐ | — | — | ◐ | ✓ |
| Engineering and non-engineering work | ✓ | ◐ | ◐ | ✓ | ✓ |
| Mix harnesses and models as needed | ◐ | — | ✓ | — | ✓ |
| Acceptance criteria built into requests, sign-off on evidence | — | — | ◐ | — | ✓ |
| Plugin review, publishing and scoped access | ◐ | — | — | ◐ | ✓ |
| Enterprise-grade production and team collaboration | ◐ | ✓ | ◐ | ◐ | ✓ |
| Ready to use without building a business layer | — | ✓ | — | ✓ | ✓ |
| Hosted service | ✓ | ◐ | ◐ | ✓ | ◐ Self-hosted |
✓ Supported ◐ Partial or varies by product — Not supported. Based on how each category typically works.
The hard parts, handled:
| Hard part | How SoduWork handles it |
|---|---|
| Clients disagreeing about facts | The Hub is the single source of truth: issues, workflows, approvals and evidence live in the Hub, so every client sees the same facts |
| Agents getting access they shouldn't have | Access is granted, not assumed: organization policy, runtime availability and the user's choices together decide what a session can use. Installing something doesn't authorize it |
| Lock-in to one harness | One adapter interface for every harness: higher layers never depend on a particular harness. Switch harnesses mid-conversation and keep the history and working folder |
| Methods that can't move | Method and environment are separate: plugins define the method; harness, model, runner and secrets are chosen when the plugin is bound. Plugin versions are immutable and portable |
| Work nobody can trace | Interaction and execution are separate: the client handles chat and approvals; the runtime and runners do the work — locally or remotely — and every run lands in the audit record |


The runner connects to the Hub over a WebSocket. Every dispatch has an ID, so the same task is never taken twice. Each harness runs in its own subprocess. Events are saved on the execution host before they are sent, and are sent again after a disconnect.
Telemetry and updates. This public source does not include a separate analytics or crash-reporting upload to MDiRi. The client defaults to https://service.soduwork.ai; its operational requests go to the Hub selected under Connection on the sign-in screen. You can change this address, and your saved selection is kept across restarts and upgrades. Using the hosted service sends those requests to its operator. Selecting your own Hub keeps this traffic with your deployment.
Not if any of these hold: you re-check AI output yourself, you can't reproduce last month's perfect workflow, or you just want your AI tools to share one standard. Solo use needs no team setup — run it locally with SQLite in one command, and the organization features simply stay unused until you need them. Same plugins, same workbench, more people when you want them.
SoduWork doesn't replace agent harnesses — it runs them inside a team workspace and adds the layer they don't have: one shared issue pool from request to sign-off, acceptance criteria built into every issue, an independent AI review step, plugin review with scoped grants, and an audit trail. People set the steps, checkpoints and approvals.
It's an active public beta. Every release has release notes, and the release pipeline runs repository checks and test suites before a version is published. See the Roadmap for what's next, and Contributing if you want to join.
Accounts underpin organizations, permissions, collaboration and audit. When you self-host, you sign in to your own Hub and your data is stored there. See Where does my data go?
More answers are in the FAQ on the developer site.
| Status | Item |
|---|---|
| ✅ Shipped | Cloud runner (Personal Agent) support |
| In progress | Development kits for plugins and apps |
| In progress | Sandbox management for cloud runners |
| In progress | Web and mobile apps |
| Planned | Marketplace support for apps and plugins in enterprise and private deployments |
Ideas are welcome in Discussions.
Issues, ideas, plugins, docs and code are all welcome.
| Goal | Read |
|---|---|
| Build a plugin | Plugin guide |
| Build an app | Apps |
| Add a harness | Add an ACP harness |
| Deploy and operate | Deploy for a team |
| Commands and configuration | Reference |
Please don't report security issues in public issues. Follow SECURITY.md to report them privately.
SoduWork's own code is copyright MDiRi and released under the MIT License. Third-party code, fonts, graphics and bundled tools keep their own licenses; see Third-party notices. The SoduWork name and logo are trademarks of MDiRi and are not licensed under MIT.
Thanks to Electron, React, FastAPI, SQLAlchemy and the wider open-source ecosystem, and to the agent projects SoduWork runs. Being listed here doesn't mean a project endorses SoduWork.