SoduWork ends a request when the person who asked signs off, not when the chat stops

The README opens with a sentence that is also the product spec: a chat ends with an answer, a SoduWork request ends when the person who asked signs off.
That is a small reframing with large consequences. If a request only closes on sign-off, then "done" has to be defined before the work begins, somebody other than the author has to check it, and the whole thing has to leave a record. SoduWork, from mdiri-ai, builds those three requirements into one shared pool of issues. It reached 6,170 stars after a 29 October 2026 launch and is in public beta.
What a request looks like
You raise a request with a goal, the inputs, and the acceptance criteria. An AI pass can refine the wording and add criteria you missed. Then you assign it, to yourself, a colleague, or an agent, and the assignee's harness runs it.
When the work comes back, a separate AI review checks each criterion. If something fails, the work returns with reasons attached. If it passes, the assignee confirms, and the requester signs off. Only then is the request closed.
The diagram in the repo has six boxes and one dotted line: raise, refine, assign, run, review, sign off, then save as a plugin, and the dotted line goes back to run because the plugin is reused.
One harness works, another reviews
The harness is the agent runtime, not the model. SoduWork connects nine of them through one adapter interface: Pi, OpenCode, DeepSeek Harness, Qwen Code, Kimi Code, Codex Harness, Codex, and two more behind catalog declarations. Programs that speak the Agent Client Protocol can be added by describing them.
The design consequence is that the reviewer does not have to be the author. Pick Pi with one model to do the work, then have OpenCode with a different model review it against the criteria. Two different harnesses, two different failure modes, and a review that is not the model grading its own homework.
You can also switch harnesses mid-conversation and keep the history and working folder. That sounds like a convenience feature until you have spent a week with three tools and a drawer of subscriptions; then it is the reason to use one workbench.
From one good run to a team asset
The second half of the loop is the part teams usually skip. When you and the AI get a task right, you save the conversation as a plugin. The plugin goes through review, gets granted to a team or a person, and can be revoked. Anyone on the team can run it and contribute improvements.
The pitch is that knowledge stops dying in a chat thread. The realistic version: a plugin is a saved method, and a saved method is only as good as its review. SoduWork makes review a gate rather than a suggestion, which is more than most plugin systems do.
Governance sits underneath all of it: roles, usage, cost, and an audit log, self-hosted. Who asked, who did the work, which plugin ran, who approved it. If you are the person who has to answer "why did the agent send that email", this is the table you want.
Running it
From source, the prerequisites are Python 3.12.x and Node 22.x (22.19.0 or later in that series), plus Git. Clone the repo and start it:
sh start.sh --desktop # macOS / Linux
.\start.cmd --desktop # Windows PowerShell
First run installs dependencies, starts the Hub, the admin console and the desktop client on the same machine. You set a password for the admin account, 12 to 128 characters, then add a model provider under Model catalog.
That is the step people get wrong, and the README calls it out: saving a service credential does not create a model. You have to open the service you connected, choose Add model, paste the provider's exact model ID, and mark it as a chat model. Pi uses the SoduWork catalog; Codex uses its own account and needs its own sign-in, and will not inherit the API key you just added.
For a team, Docker does the whole stack:
python3 ops/self-host/deploy.py up
python3 ops/self-host/deploy.py credentials
That brings up the Hub, the admin console and PostgreSQL on one server. Put an HTTPS reverse proxy in front and hand out the address. Desktop clients for Windows, macOS and Linux are on Releases, and the sign-in screen has a Connection field where you paste the Hub URL.
There is a hosted demo at service.soduwork.ai if you want to look before installing. The README is explicit that it is for evaluation only and that you should not upload real business data.
Where it will strain
Four things to weigh.
The review is still a language model. "A separate AI review checks each one" is a real improvement over self-assessment, and it is not verification. A criterion like "the export contains all 47 rows" is checkable; a criterion like "the tone matches our brand" is a model's opinion with a checklist taped to it. The quality of the whole loop is the quality of your acceptance criteria, which is a writing skill, not a product feature.
Nine harnesses is a support burden. Adapters rot. The harness with the fastest release cadence will break first, and the parts that depend on it (plugins, review, history) break with it. A small team can keep two or three adapters honest and will stop testing the rest without noticing.
There is also a category problem. SoduWork is competing with project trackers, agent orchestrators and chat clients at the same time, and each of those has a decade of expectations attached. Filing a request in SoduWork means not filing it in Linear, and the agent may be the only participant who prefers the new tool.
Beta means beta. Public beta, evaluate before production, no compliance certifications mentioned. Self-hosting is the right answer for teams that care about data residency, and self-hosting also means you own the upgrades.
Adoption is the actual gate. A request pool only works if people file requests in it. Ten engineers with ten private chat windows will route around any workbench, and the audit log will be empty and accurate.
The idea worth stealing
Set the bar before the work starts. Have somebody else check it. Make a person sign off. Save what worked.
None of those four are new. They are how agencies, editors and QA departments have always worked, and they are exactly what chat interfaces removed. SoduWork puts them back and lets you choose which agent runtime does the typing.
If you run a small team and the recurring complaint is "the agent said it was done and it wasn't", the acceptance-criteria field is the cheapest part of this to copy, with or without the software.
Related articles
Six Amateurs, 1,113 Episodes, a Thousand Yuan of Compute an Episode: How an AI Drama Got Hand-Built
The first ten seasons of Armed Into Song run to roughly a thousand episodes and have passed ten billion cumulative views, with six core creators and not one of them film-school trained. Compute runs 1,000–2,000 yuan an episode; a normal shot takes three rolls, a battle twenty-plus, and anything that looks flat is graded a sixty and thrown out.
One server, one AI streamer: inside the code that runs a nightly livestream
open-yoki is the production code behind an AI host who streams nightly on Binance Square, with the persona, voice and money features stripped out. Headless Chrome draws the frame, ffmpeg pushes RTMP, a folder of Markdown decides what she feels and what she posts.
Pinky lets an agent drive your phone, and holds the Send button until you click Allow
Pinky connects any MCP agent to a real iPhone or Android phone and enforces approval in its own gateway rather than in a prompt. Banking apps are invisible, password fields are off limits, and on Android 13 the agent gets a hidden second screen.
ByteDance's SeedRealtime takes two frames a second and decides when to speak
Doubao-Seed-Realtime-1.0 folds camera frames, screen frames and speech into one full-duplex session, at no more than two frames per second. The interesting decisions are the frame rate, the proactive speaking, and the audit reply that ships switched on.