SoduWork Turns AI Chats into Signed-Off Work Orders: A Hands-On Test
The likely breakout of 2027: AI workbenches
A take: AI workbenches are the likely breakout of 2027. Not the kind of chat box that exchanges a few pleasantries, but a tool that can carry a task from receiving an order to acceptance, and requires someone to sign off. I just ran SoduWork locally over the past couple of days, cleared the three hurdles of China's network, got it installed, running, and it does have something to show. Let me use it as an example to explain how far this class of tools has come.
Test environment: Windows 11 · PowerShell Tested object: mdiri-ai/SoduWork, MIT license
Below follows the order of my operations: first understand what it is, then clone the code, prepare the environment, start it, and finally say what it looks like when running and whether it's worth installing.
一、First, figure out what this thing is
The official self-description is "enterprise-grade AI-native engine" and "digital AI employee, with digital methodology first." In plain words: it's an AI work order system. You submit a request ticket, write down what you want and what the acceptance criteria are, hand it to the AI to execute, and a human signs off when done. A chat box ends with a reply; this one ends only with sign-off.
The first line of the README says it all:
"A chat ends with an answer. A SoduWork request ends when the person who asked signs off."
A few terms keep showing up; let's go through them:
| Concept | Official name | My understanding |
|---|---|---|
| Harness | execution base | The hands that actually do the work. Pi, Codex, OpenCode, Qwen Code are agent runtimes, not models, and can be swapped anytime; chat history and working directory are kept |
| Model | Model | The brain, connected via your own API key. Hand and brain are separate |
| Request | Request | A work order, not a conversation: goal, materials, acceptance criteria, written before work starts |
| Plugin | Plugin | A successful one-off run gets hardened into a versioned capability package; after review the whole team can use it |
| Runner | execution machine | The workstation that does the work, local or on your own cloud server, connected to the Hub via WebSocket |
It closes the loop of "request → assignment → AI execution → AI self-check → human acceptance → plugin hardening"; the management side with audit logs, role permissions, execution network, usage cost is indeed rare among open-source AI tools.
Background data (as of 2026-11-17, I verified with the GitHub API): the repo was open-sourced at the end of October, 1,299 stars and 182 forks in three weeks, README explicitly marked as public beta. The tech stack is Electron + React client, FastAPI + SQLAlchemy backend, React/Vite admin panel; single-machine uses SQLite, team deployment uses PostgreSQL 16.
One small detail shows how fast it's moving: the README table header says "9 supported Harnesses" but only lists 7. Documentation is running faster than the table. That flaw is noted here; the "shortcomings" section mentions it again.
二、Getting it running: from cloning to login page
First, which path to take. It offers three installation methods:
- Source launch:
\\.\start.cmd --desktop, fast experience, single machine, requires Python 3.12+ and Node 22.19+, no Docker, Git, or hand-written.env; - Docker self-hosting:
py ops/self-host/deploy.py up, for team server deployment, spins up PostgreSQL 16; - Download the installer and connect to the official cloud, register an account at
service.soduwork.aito try.
I took the first path. It's the easiest: one command brings up the Hub (backend service), the browser admin panel (127.0.0.1:18989), and the Electron desktop client; all data stays in local SQLite, no touching the existing environment. I also read the documentation for the second path, which is detailed, auto-generating database passwords, admin passwords, signature keys, and only exposing the address after health checks pass. But my machine's Docker engine won't start (Docker Desktop reports a launcher lock, my local issue, not the project's), so the Docker route was only verified against the docs, not tested. I'm saying this upfront so you don't think I validated it.
Start cloning. The first attempt hit a wall:
fatal: unable to access 'https://github.com/mdiri-ai/SoduWork.git/':
Recv failure: Connection was reset
Expected. The repo has nearly 4,900 files; using an acceleration prefix, it took 30 seconds:
git clone --depth 1 https://gh-proxy.com/https://github.com/mdiri-ai/SoduWork.git
(ghfast.top timed out this time, gh-proxy.com worked. These mirror sites' availability drifts, use whichever responds.)
start.cmd looks like one command but downloads three big things in the background. Without handling China's network, you'll likely get stuck at step three and be confused.
Python packages (two dependency sets for Hub and Runner):
$env:PIP_INDEX_URL = 'https://pypi.tuna.tsinghua.edu.cn/simple'
Electron binary (100 MB level, needed by desktop client, default downloads from GitHub):
$env:ELECTRON_MIRROR = 'https://npmmirror.com/mirrors/electron/'
This mirror worked great: 67.6 seconds to download.
PortableGit (~58 MB). Windows desktop startup prepares a built-in command-line environment (Bash + Git) for the agent, pinning a specific version. My most reliable method is to download it to local from npmmirror, verify the official SHA-256, then feed the path to the launcher:
curl.exe -sSL -o PortableGit-2.55.0.4-64-bit.7z.exe `
https://registry.npmmirror.com/-/binary/git-for-windows/v2.55.0.windows.4/PortableGit-2.55.0.4-64-bit.7z.exe
# Check: should be 016e84230a3767f0c6b3788e79ba0c58a17377086801719d46700fca4f7b36b5
Get-FileHash .\PortableGit-2.55.0.4-64-bit.7z.exe -Algorithm SHA256
$env:SODUWORK_PORTABLE_GIT_ARCHIVE = 'C:\full\path\PortableGit-2.55.0.4-64-bit.7z.exe'
The launcher does a fixed digest check on downloaded items, only allowing the mirror content to match official bytes, a point I liked; even a poisoned mirror can't fool it. It took 41.1 seconds to prepare and unpack.
Three environment variables plus one local archive are everything you need to know for deploying SoduWork in China. Then launch:
.\start.cmd --desktop --no-browser --password-stdin # automation usage
.\start.cmd --desktop # normal usage, opens browser + client
The first run asks you to set a password for admin, 12–128 characters, input is not echoed, no asterisks, type then press Enter, don't think it's stuck. This password exists only in the local instance, not any online account.
The launcher splits the process into 4 steps, actual measured times:
| Step | Content | Measured time |
|---|---|---|
| 1/4 | Create isolated Python environment (venv) | 12.7 s |
| 2/4 | Install server dependencies (FastAPI etc.) | about 2 min |
| 3/4 | Install admin dependencies + build (Vite) | about 1.5 min |
| 4/4 | Desktop client preparation: Electron download | 67.6 s |
| 4/4 | Desktop client preparation: CLI environment (PortableGit) | 41.1 s |
| — | Start services + database migration | about 10 s |
Total about 5.5 minutes. (Second start reuses dependencies, takes tens of seconds.) When ready, the terminal prints:
SoduWork is ready: http://127.0.0.1:18989
Administrator: admin
The browser first shows an intro page, click "Enter admin panel" to see the login; the desktop client pops up on its own, and the login page clearly shows it's connecting to 127.0.0.1:18989:
All data lives in this directory; deleting the database means deleting this, backing it up means backing this up:
| Content | Location (Windows) |
|---|---|
| Source-mode data directory | %LOCALAPPDATA%\SoduWork\source-service |
| Business database | Same, soduwork.db (SQLite) |
| Startup log | Same, startup.log (troubleshooting relies on it) |
| Instance key | Same, local-service.json (don't share) |
三、What tripped me during startup
The first start wasn't smooth. Step 2 failed directly, one-line log:
ERROR: Can not perform a '--user' install.
User site-packages are not visible in this virtualenv.
The cause isn't SoduWork but the environment preset PIP_USER=1 (some Python distributions or toolchains set it). Adding --user inside a virtualenv always fails:
Remove-Item Env:PIP_USER -ErrorAction SilentlyContinue
# or set $env:PIP_USER = '0' before startup
I noted this: any domestic AI project that reports a --user error while installing pip in a venv is probably this.
四、What it looks like when running
The admin panel doesn't look like an open-source demo; more like something made by people who've done enterprise software. After login, a full left-hand navigation column has governance, operations, monitoring, permissions, audit sections.
The homepage's four cards are management-view: request pool, request delivery rate, pending risks, online rate (1/1 online). Below are "department demand status" and "what to handle now (overdue requests / execution anomalies)", all indicators for managers, not logs for engineers.
The model directory page is the best-designed. It has 10 built-in providers, major ones like Tongyi, Claude, DeepSeek, each card marks "missing key/missing address/configured", with real-time load: generating, queue waiting, concurrency limit, average queue time, HTTP 429 count.
The audit log is even more interesting. It recorded every step of my configuration changes: time, operator, action type, object, details, all present.
Who changed which model service at what time, that kind of record is what distances it from "personal AI tools."
The desktop client is restrained. After login, left side has new chat/projects/recent chats, middle has a one-line hint and an input box:
The small row under the input box carries a lot of information: permission scope (default "full access", can change), working directory (default workspace), session ID, model selector, Harness selector (Pi). That means within the same session, both model and Harness can be swapped, records and directory remain, which is what it calls "hand-brain separation."
How does real work go? I used the test question from the official docs, the one that best shows whether the AI obeys:
Convert the following simulated meeting minutes into a to-do list, listing the responsible person and deadline; for missing information write "to be confirmed": Xiao Wang is responsible for organizing requirements, to be done by Friday; Xiao Li checks the quote.
It returned in 3 to 6 seconds, and strictly followed:
| # | To-do | Owner | Deadline | Remark |
|---|---|---|---|---|
| 1 | Organize requirements | Xiao Wang | Friday (specific date to be confirmed) | — |
| 2 | Check quote | Xiao Li | To be confirmed | — |
Below that it added a note: only "Friday" was mentioned without a specific date, so it was marked to be confirmed; Xiao Li's line had no deadline, written as to be confirmed; if you want to add priority or deliverable, you can also list as to be confirmed.
This step is actually key. Many models would invent a specific date for "Friday," or simply drop the "to be confirmed." This one didn't. Whether it obeys instructions and marks its own uncertainty is the line between "can do real work" and "only for fun."
I also asked it "what model are you?" and it honestly said: underlying GLM-5.3-Flash, working through the pi encoding agent framework, can read/write files, execute commands, operate a browser.
A few details were comfortable to use. "Restore this round" when model calls fail: context and tool results are kept, fix the config and click to resume, no need to re-explain the request, a design that understands real work. Process is visible: above replies it says "took 12s / deep thinking / ran command," not a black box. Permissions are visible too: beside the input box there's a persistent "full access" hint, click to change, doesn't secretly grab all-disk permissions.
Shortcomings to note: model capabilities default to "unconfirmed," image input, tool calls, streaming output need to be checked manually or by clicking "detect capabilities," easy to miss on first connection, which shows as "looks configured but doesn't work right." You must bring your own API key; no key, it won't even answer the first sentence, which is by design (it doesn't sell models), but unfriendly to pure beginners; official online trial environment service.soduwork.ai is a compromise.
There are also maturity-related points. Docs and implementation have a small mismatch, the earlier 9 Harnesses header with a 7-row table. Docker wasn't tested this time (local Docker engine failure), the docs say auto-generated passwords, health checks, backup fallback, looks complete, but untested doesn't mean works. It's still in beta, README says "evaluate thoroughly before using in production," don't put critical business on it bareback.
One last point about security: organizational keys are stored in plaintext in the database. The admin panel says "keys are only stored in the backend database, no API returns plaintext," which is true, the API doesn't display them. But in the local SQLite, I saw plaintext in model_provider.api_key. For self-hosting, you must rely on disk encryption plus directory permissions, and don't put the data directory on a shared drive.
五、Is it worth deploying?
Split by audience:
| Who you are | Verdict | Reason |
|---|---|---|
| Individual developer with API key | Worth installing, half a day to get running | One command brings up the whole stack, request pool plus plugin hardening is practical for solo development or side gigs, models can be swapped to cheaper ones |
| Team of 3–20 people | Worth evaluating | Audit logs, role permissions, execution network, usage cost are really implemented, not decorations; plugin review and distribution can turn one person's specialty into team asset |
| Enterprise / needs private deployment | Do POC first, don't rush to production | Positioning is right (self-hosting, governance, audit), but still public beta, mobile/web end on the roadmap |
| Just looking for a free ChatGPT shell | Don't install | It has no models by default, must bring your own key; value is in process, not chat |
In one sentence: SoduWork's value is in process, making AI output verifiable, reusable, and accountable. For individuals it's efficiency, for teams it's order. If you're stuck at "AI works great for me, but the team can't use it," spend half a day trying it.
Appendix A: Command cheat sheet
# 1. Clone (China mirror)
git clone --depth 1 https://gh-proxy.com/https://github.com/mdiri-ai/SoduWork.git
cd SoduWork
# 2. Network acceleration (PowerShell, current session)
Remove-Item Env:PIP_USER -ErrorAction SilentlyContinue # prevent venv install failure
$env:PIP_INDEX_URL = 'https://pypi.tuna.tsinghua.edu.cn/simple'
$env:ELECTRON_MIRROR = 'https://npmmirror.com/mirrors/electron/'
$env:npm_config_registry = 'https://registry.npmmirror.com'
# PortableGit either mirror root or local archive
$env:SODUWORK_PORTABLE_GIT_MIRROR = 'https://registry.npmmirror.com/-/binary/git-for-windows'
$env:SODUWORK_PORTABLE_GIT_ARCHIVE = 'C:\...\PortableGit-2.55.0.4-64-bit.7z.exe'
# 3. Launch
.\start.cmd --desktop # Hub + admin + desktop client
.\start.cmd # only browser admin
.\start.cmd --check --desktop # health check only, no dependencies
.\start.cmd --desktop --port 18990 # change port if occupied
# 4. Team Docker self-host (not tested this time)
py ops/self-host/deploy.py up
py ops/self-host/deploy.py credentials # view initial admin password
Appendix B: References
- Repo: https://github.com/mdiri-ai/SoduWork
- Installation and startup: https://github.com/mdiri-ai/SoduWork/blob/main/docs/getting-started.md
- Docker self-host: https://github.com/mdiri-ai/SoduWork/blob/main/ops/self-host/README.zh-CN.md
- Official site: https://soduwork.ai · Online docs: https://dev.soduwork.ai/zh-CN/
All screenshots taken from real local runs. The project is in public beta; UI and steps may change with versions, refer to official docs.