How Wang works

You onboard Wang as your AI employee — your CTO hire. Tell it what you need in plain words and it builds and runs your Wok for you: it sets up the database, auth, storage, realtime, functions, scheduler, email, SMS, hosting, the built-in AI model, and key management, then builds your app on top, checks each part actually works, and gives you something live. You own what comes out — the features, the code, the look of the software that runs your business. Wang builds against a live Wok throughout the process, so what you get back has already been provisioned, deployed, and tested on the real system you'll use.

How Wang builds your Wok

Every change — whether you asked for it or Wang did it — goes through the same four steps on your real Wok:

1. Work on the real thing — Wang changes your actual Wok, not a draft or a guess.
2. Build it — sets up the data, the rules, the automations and the app.
3. Try it — Wang actually uses what it just built: for a UI, it opens your app in a real browser and reads the console and network requests for errors a plain status-200 check would miss; for data and APIs, it runs the queries and calls and checks the responses.
4. Publish and finish — if something's wrong, Wang fixes the cause; if Preview passes, the normal release gate promotes it and Wang hands you the verified Published link. A Preview link is returned only when you asked to review Preview or publishing is blocked.

Because Wang tests on the real, running Wok, the classic "it worked while I was building it but broke when it went live" problem doesn't happen.

What Wang can actually do

Ask in plain words and Wang takes the action — set up a Wok, change your database, deploy code, schedule recurring jobs, add logins, register and wire up a domain, put your site online, send and read email, search what you've taught it, top up credits, and connect Slack or Telegram. Developers can drive the same actions directly over the API.

Inside an app workspace, Split opens the running app beside Wang. Desktop keeps a 1280-pixel layout viewport and scales it to fit, while Mobile uses a phone viewport. Starting Annotate gives the preview the full detail area so individual elements are large enough to select.

Teach Wang about your business

Onboarding is a conversation: Wang wants to understand where your business is now, where you want it to go, and the end goal. Every current plan can chat with Wang directly in the dashboard; Free uses DeepSeek Flash and its included monthly model allowance. What you teach Wang applies wherever it runs:

• Instructions — a short set of rules Wang always follows: tone, naming conventions, "always provision in ca-central-1", "confirm with me before buying any domain", "we bill in CAD". They're layered on top of Wang's base behavior (and never override the safety rules), so every chat acts like it already works at your company.
• Knowledge — keep a small set of plain-text facts that Wang itself should remember: your menu, operating policy, naming conventions, or product details. Wang searches them on each question (typo-tolerant) and answers from your facts instead of generic defaults.
• App Documents — when an app needs a large, governed corpus, open its Documents resource. TellWang spins up LangChain or LlamaIndex on request, vectorizes large batches, applies principal/role/group/deny rules before anything reaches a model, and lets you build or turn off a support agent whose answer claims are checked against their cited passages, create new files from selected sources, or analyze the collection. A file can be blocked for one customer-facing agent while remaining available for internal generation and insights. Generated files remain in Documents after refresh. A hard “Do not use” action removes an indexed file from all three uses.
• Analytics — open Audience analytics for aggregate country and region, device, language, campaign and source, privacy-safe clicked-element labels, conversions, top pages, and daily traffic. Session replay is the adjacent submenu: enable masked or full-fidelity recording with an owner privacy acknowledgement, filter Published sessions to 24 hours, three days, or seven days, then triage entry pages, duration, client errors, rage clicks, and dead clicks before opening playback. Aggregate analytics is cookieless; visitor identifiers rotate daily, and raw IP addresses, age, and gender are not stored or inferred. Replay remains owner-controlled, retention-bounded, off by default, and disabled for browsers advertising Global Privacy Control or automation.
• Design system — upload your brand as files: a Claude Design export, a tokens.json, a theme.css, a brand guide. Wang reads them when it builds and matches your colors, type and spacing instead of a tasteful default — and a stylesheet of your tokens is auto-served at /assets/tokens.css in every Wok, so generated pages reference your brand variables directly.

Your instructions and knowledge base apply everywhere Wang runs — the dashboard chat and the Slack/Telegram channels — so the whole team gets the same context-aware Wang; your design system applies whenever it builds a frontend. Chats live at the organization level rather than inside one Wok. A focused Wok is an optional starting point, and Wang can still work across several Woks because each action names and authorizes its target explicitly. Pro and Scale organizations run through their own OpenCode Agent Cell. Each active chat gets an isolated worker whose access is bound to the human who started it, while TellWang keeps identity, approvals, tools, audit, and the conversation in one place.

Picks up where you left off

App building in Wang uses a coding model. Free organizations can select DeepSeek Flash; paid organizations can also select DeepSeek Pro or Kimi, and any member can connect ChatGPT or Grok. The composer keeps the message unsent until a coding model is selected. The server also removes build authority from a Private Wang Agent Cell's scoped runtime key, so its internal MCP calls, TellWang CLI commands, and Git pushes cannot bypass the rule through a generic chat or stale client. Private remains available to technical builders through the organization model gateway.

While Wang works, the chat shows verified lifecycle state and a safe ledger of the files, releases, and checks its tools touched. OpenCode reasoning appears in a collapsed Thinking accordion; public model text appears as ordinary Wang messages outside it. If TellWang's completion evaluator asks the same Agent Cell to continue, that internal guidance also stays in Thinking instead of appearing as a customer message or public narration. A server-validated lollipop finish-run decision prevents completion from forcing another turn when only managed DNS propagation remains; it cannot bypass ordinary incomplete work. The working stopwatch remains visible beside the run. Existing conversations keep their historical mixed rows in Thinking because content recorded before typed streams cannot be split reliably after the fact. Each independently accepted final answer stays a separate immutable message. Opening a chat under a project opens its app Preview beside the conversation automatically; closing Preview still leaves it closed until another project chat is opened. The capability matrix above the Agent Cell composer attaches the selected capability behind the scenes for app creation, domains, email, RAG, functions, automations, diagnosis, and other common work. The composer and durable user message stay exactly as the user wrote them. Click an active capability again to deselect it. The composer accepts one PNG, JPEG, or WebP screenshot through its upload control or a clipboard paste. A compact cue beside Send shows Ctrl ↵ on Windows and Linux or ⌘ ↵ on macOS; plain Enter remains available for multiline messages. Wang chat's model row has one DeepSeek pill: Free can select Flash, while paid plans also see Pro. It also lists paid-plan Kimi, personal ChatGPT, and Grok Build through a member's Grok or SuperGrok subscription; Private is not offered in that row. Both personal connections use device sign-in; TellWang never asks for an xAI API key. For Wok-focused work, TellWang prepares the named app's retained Preview checkout before OpenCode starts. Static files follow the browser route: frontend/index.html is /, while frontend/<route>/index.html is /<route>/, so two index files are separate pages rather than duplicate checkouts. Root-page requests start at the root file; site-wide work also inspects shared assets and every routed page that owns the repeated UI. OpenCode follows that checkout's AGENTS.md and TELLWANG.md. TellWang interrupts repeated inspection without editing, verification, or a concrete blocker, and independently checks the final answer against the recent user goal and tool evidence before marking the run complete. The conversation, queue, active run, typed model streams, safe progress, and execution stopwatch survive a refresh. Team members can submit attributed follow-ups while Wang is busy and share a private group chat without filing a cross-Wok task under the first app it touches.

Managed DNS propagation is an expected external wait, not abandoned work. Once TellWang has published the records and the required one-shot verification reports a future retry time, Wang can finish with a truthful pending-activation status and no stopped-before-finishing warning. A domain whose external DNS still needs customer action remains incomplete.

One Wang, every channel

Connect a Telegram or Slack bot through the channel API so your team can message Wang from the chat app they already use. Engineers drive the same Wok over MCP or the REST API; the dashboard shows the connection address and prepares a scoped prompt.

Confirmation & guardrails

Wang handles safe, undoable work on its own, but asks first before anything risky — deleting data, buying a domain, sending email, spending money. And it can only ever touch your team's own Wok — never another team's, even if someone tries to trick it into it.

Portable by default

A Wok is yours to take. Your data is plain Postgres you can pg_dump on any plan, so it's always yours. On Pro and Scale, a Wok's entire definition — schema, policies, functions, env, storage config, scheduled jobs, and infra shape — also exports to Terraform HCL in one call:

curl https://tellwang.com/v1/woks/$WOK/terraform \
  -H "Authorization: Bearer $TELLWANG_KEY" > wok.tf

terraform init && terraform apply
# → redeploys this Wok on a different cloud, your own account, or self-hosted.

No lock-in: TellWang earns its keep by running your Wok well, not by trapping it.

Proven in practice

We built a full point-of-sale system this way first, and building on the live system beat writing code and deploying it separately every time. TellWang is that, packaged for your team.