SoduWork – Open-source workbench where agents must prove their work SoduWork, an open-source, self-hosted multi-agent workbench from mdiri-ai, entered public beta with a workflow in which one harness performs a task and a second harness on a different model reviews it against pre-set acceptance criteria, with a human requester signing off before a request is considered done. The workbench supports Pi, OpenCode, DeepSeek Harness, Qwen Code, Kimi Code and Codex Harness, requires Python 3.12+, Node.js 22.19+ and Git, and starts via `sh start.sh --desktop` on macOS/Linux or `start.cmd --desktop` on Windows. The project's documentation warns that a successful connection probe does not guarantee successful inference, since balance, model permissions and protocol support still matter. English · 简体中文 https://github.com/mdiri-ai/SoduWork/blob/main/README.zh-CN.md 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 https://soduwork.ai | Live demo https://soduwork.ai | Documentation https://github.com/mdiri-ai/SoduWork/blob/main/docs/README.md | Docs website https://dev.soduwork.ai | Releases https://github.com/mdiri-ai/SoduWork/releases | Discord https://discord.gg/AXYNrAYHy | 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. 1. Add a chat model as well. Under Capabilities → Model catalog , find the service you connected, choose Add model , enter the provider's exact model ID, select Chat model , and save. Saving a service credential alone does not create a selectable model. 2. Open the desktop client and sign in to the same Hub. Use --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. 3. Choose New conversation, Pi, and the model you added. Pi uses the SoduWork model catalog. Codex uses its own account and needs its own sign-in; it does not automatically use this API key. 4. Try a small task. For example: “Turn these fictional meeting notes into an action table with owner and due date. Mark missing information as unconfirmed: Alex will organize the requirements by Friday; Sam will check the quotation.” Check the returned table against the notes. 5. Create a requirement when you need assignment and acceptance. Describe the goal, inputs and acceptance criteria, open Arrange execution , choose an Agent and model, and inspect the delivery before accepting it. A normal conversation does not require a project, plugin or automation first. 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 https://github.com/mdiri-ai/SoduWork/blob/main/docs/getting-started.en.md send-your-first-ai-message for field-by-field setup and troubleshooting. Try the live demo Download the client from Releases https://github.com/mdiri-ai/SoduWork/releases/latest , 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. Install the desktop client Follow the illustrated desktop installation and first conversation guide https://github.com/mdiri-ai/SoduWork/blob/main/docs/desktop-quickstart.en.md . | OS | Package | |---|---| | Windows | SoduWork-