Ciliclaw Digital Workforce · Developer preview
Send a message in DingTalk, or say it on the desktop. The agent works out what you mean, breaks it into steps, calls the tools, runs them to the end, and brings the finished work back to you.
Documents, spreadsheets, decks, code, meeting notes, research — none of it needs watching. You receive the result, then review it.
Turn last quarter’s East China sales into a monthly review deck — I need it for tomorrow.
Sales ledger · 2026 Q3 · East China
Quarterly analysis.xlsx · 4 charts
East China Q3 review.pptx · 12 slides
With a note on how figures were derived
What it takes on
Not a chat box you have to steer. Each of these it does end to end — you give a sentence, you get a finished thing.
Writing, spreadsheets, slides. It finds the data, works out the numbers, lays it out, and checks its own formatting and figures before handing it over.
Turn the survey responses into an analysis report, with charts.
Reads the codebase, locates the problem, writes the fix, runs the tests. Work is dispatched to the coding agent already installed on your machine — the code never leaves it.
Add phone-number login to the user module, with tests.
Transcribes, separates speakers, pulls out decisions and actions, drafts the write-up, and matches attendees against the directory.
Write up the standup and assign the action items.
Browses under control, extracts what matters, cross-checks it, and returns a written summary with sources — not a pile of links.
Compare these three suppliers on credentials and recent news.
Answers from your own policy documents, citing the source and version on every claim. Expired or unreviewed material is never used in an answer.
What is the lodging allowance? How does it work for long postings?
Bulk renaming, filing, format conversion, content extraction. Every local file action leaves a record you can go back to.
File these contracts by signature date and rename them consistently.
Where you use it
DingTalk is the main way in: you are already there, so hand the work over there. When you want to watch the work happen, the desktop app takes over.
Best for · Starting from anywhere
Send one message in a thread you are already in, and the work begins
Best for · Watching the work and local files
See each step as it happens and redirect it mid-run
How it works
“Automatic” is not throwing your request at a model and hoping. There are four fixed steps — understand, plan, run, hand over — and each leaves a record you can read.
A sentence becomes a stated goal: what to produce, for whom, by when, and from which source. When something is genuinely missing it asks once rather than guessing onward — one question beats three wrong attempts.
The goal becomes a list of concrete actions, each bound to a real tool: what to look up, what to compute, what to produce, where to send it. The plan is written down, so you can see how it intends to proceed and change it mid-run.
It works through the plan without you. A failing tool is retried or routed around; it only comes back to you when it genuinely cannot proceed. Actions that change something — writing files, running commands — confirm first, by the rules you set.
The finished work returns to where you asked, along with what it did, which data it used, and anything it was unsure about. You do one thing: look, then decide whether to use it. This is the only part of the run that costs you time.
Security & deployment
Work is executed by the local unit in your own environment. Files, code, and data stay on your network. Paired with a local model, even the inference request never leaves.
Documents, code, knowledge bases, and conversation records all live in the company’s own environment. With a local model, inference stays inside too — there is no path that sends company material to an external model service.
Work is picked up and completed by the local execution unit in your environment. File reads and writes, commands, and scripts each leave a record tied to the person, the task, and what authorised it.
Actions with real effects — writing files, running commands, sending messages — ask for confirmation by the rules you set. Which run freely and which must ask is configured centrally.
Data, pages, and statistics are scoped to one company; nothing is shared or joined across two. Our people cannot see your content by default, and assisting with an investigation requires a recorded, time-boxed grant.
Deployment
Every component runs on your own servers or private cloud, local model included.
Data and execution stay internal; model calls go to a service you have approved.
EVERY CAPABILITY IS A PLUGIN
Every capability is a separate plugin: office documents, the coding channel, meetings, browsing, voice, local models — each can be added, removed, or replaced on its own.
Adding a capability means adding a plugin directory and changing no central file. That is the single acceptance criterion we hold our own architecture to.
Each wires in through 11 fixed contribution points, cooperating through services and events. Third-party and built-in plugins share one mount path and one contract check.
Every registration is disposable: unmount a plugin and the tools, UI, and services it contributed disappear with it, leaving nothing behind and touching nothing else.
Built-in, organisation-published, and user-local layers stack into one runtime assembly tree. Two departments can run different capability sets with no code change and no rebuild.
Plugins
Each plugin declares what it needs, what it contributes, and how far it is trusted, validated against the manifest contract. Built-in, organisation-published, and third-party plugins all mount through the same path.
Getting live
Start with one or two of the slowest jobs, prove them, then widen. Nobody needs full coverage on day one.
Connect DingTalk and the org structure, deploy the runtime, install the first set of capability plugins.
Deliverables
Pick one frequent, slow job — usually reporting or meeting notes — and make it work end to end.
Deliverables
Widen within one department, collect what actually goes wrong in use, and fill the gaps.
Deliverables
Add agents and capabilities role by role, and settle into steady operation.
Deliverables
Request a demo
Submit an onboarding request and we will run a demo on your own scenario: you say one sentence, and watch it run end to end. Places in the technical preview are limited; teams with a concrete scenario go first.
What happens next
We use these details only to contact you. They are not used for anything else and are not shared with third parties.