SoduWork: A Self-Hosted AI Setup Where One Harness Does the Work and Another Checks It
Most AI chats end the moment you get an answer you like. But "looks good to me" isn't the same as "this is actually correct," and when a request matters, you probably want something more than a confident reply. SoduWork is an open-source, self-hosted project that rethinks that finish line: a request isn't done when the AI answers, it's done when the person who asked signs off.
What It Does
SoduWork is an AI-native engine for teams and their agents, built around multi-harness, multi-agent, and self-hosted ideas. The core loop is a separation of duties: one harness does the work, and a different harness, running on a different model, reviews it against acceptance criteria you set up front. If the review fails, the work goes back with the reasons attached.
It plays nicely with a range of harnesses—Pi, OpenCode, DeepSeek Harness, Qwen Code, Kimi Code, Codex Harness, and Codex. You can pick a harness for each step and switch mid-conversation, which means routine steps can run on cheap or local models while the important parts get a stronger model. It also supports handing a request to a colleague's agent: you file it, AI drafts the acceptance criteria, and you assign it to yourself or someone else. Their agent does the work, a person signs off, and everyone on the request can see who did what. The stack is Python 3.12+, Node.js 22.19+, and Git, and it runs on Windows, macOS, and Linux.
Why It's Cool
-
Verification is baked in, not bolted on. The idea that a separate AI review checks each acceptance criterion, with a human making the final call, is a genuinely different take on AI-assisted work. It treats AI output as something to be checked rather than trusted by default—which is exactly how you'd want to use it on anything that matters.
-
The governance story is unusually complete for a self-hosted tool. Plugins are reviewed before they're shared, granted to teams or people, and can be revoked. There are roles, usage tracking, cost visibility, and an audit log. If you're in an org that needs to control which AI does what and where, that's a lot of the hard part already handled.
-
Know-how becomes a team asset. When you and AI get a task right, you can save the chat as a plugin. After review, anyone on the team can use it and contribute improvements. That's a practical way to turn one person's working session into something the whole team benefits from—with access granted as needed.
-
Model choice is a per-step decision. Being able to switch harnesses mid-conversation and send long tasks to your own cloud runner means you're not locked into one model or one cost profile. You put the strong models where they count and keep the rest cheap or local.
-
It's honest about being in beta. The README says plainly that it's in public beta and suggests evaluating it before production use. That's the kind of candor worth noting.
How to Try It
You'll need Python 3.12+, Node.js 22.19+, and Git. Clone the repo and start it up:
git clone https://github.com/mdiri-ai/SoduWork
cd SoduWork
Then run the start command for your platform:
| Platform | Start command |
| --- | --- |
| macOS / Linux | sh start.sh --desktop |
| Windows PowerShell | .\start.cmd --desktop |
| Windows Command Prompt | start.cmd --desktop |
The first run installs dependencies and starts the Hub, the admin console, and the desktop client locally. You'll be asked to set a password for the admin account (12–128 characters). After that:
- 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, so you have something to work with immediately.
If you've already added an API key, note that you also need to add a chat model: under Capabilities → Model catalog, find the service you connected, choose Add model, enter the provider's exact model ID, select Chat model, and save.
You can find the full docs, a live demo, and the Discord link at the repository.
Final Thoughts
SoduWork is aimed at teams that want AI in their workflows without giving up control—people who care about who reviewed what, what it cost, and whether the requester actually signed off. The multi-harness review loop and the governance features are the real draw here, and the self-hosted, MIT-licensed setup makes it easy to evaluate on your own terms. Since it's still in public beta, treat it as something to kick the tires on before you build a process around it—but the core idea, that a request ends when a person signs off rather than when the model stops talking, is worth a look.
Follow @githubprojects for more developer tools and open source projects.