From your goal to a running business.
The AI team does the work. The Model Router puts the right model on every task. ZB Cloud runs it. Every cost is in the record.
Screens show Northwick Candle Co., our scripted example, at list prices.
Start with the business you want to build.
Describe it
What you sell, who it serves, and what you need next.
Confirm we got it
A short summary of your business, which you can correct before anything is proposed.
Choose the scope
The proposed systems, connections, task estimate and cost. Switch any one off and the totals follow. Choose what to build first.
The proposed build
The systems the store needs, what will be built or connected, and the estimated cost. Northwick example: a storefront, a shopping agent and eighteen supporting systems. The marker on a card means the AI team suggested it for this store.
Work with your AI team
Follow the work in shared channels. Plans, designs, code and checks arrive as they land, and you ask for changes in the same place.
The workspace
Your AI team, its tasks and its progress in one place: a plan that ticks as tasks land, a spec, a design, a change with its checks.
Check the work
The Router picks a model for each task. Checks review the output. A failed check re-routes the task to another model, and only the fix pays that price.
The Router at work
The board of twelve, each model holding the job it does best, and the tasks flowing to them. A failed check moves a task to another model; the record shows the retry and its cost.
Launch and track
Approved work is released to ZB Cloud, on the cloud chosen for that workload by price, capacity and location. The record holds every task with its model, its cloud and its cost.
How the parts work together.
Every result is checked by something that did not produce it. The workspace holds your goals, plan and progress. Inside it, AI teams do the work, the Model Router chooses the model and checks the result, and ZB Cloud runs the build.
You
Work arrives as language: a business, a feature, a fix.
Workspace
Holds your goals, the understanding, the proposal, the plan, the state of every system and the record. One business per workspace, one balance across every model.
AI teams · do the work
Eight AI teammates. The Lead holds the plan, assigns each task, checks the result and adds it to the build. A specialist is created for each task with only the context it needs.
Model Router · chooses the model and checks the result
Chooses the model for every task as it starts, from twelve on the board, chosen from 300+ AI model endpoints across 14 leading model providers, and re-routes when a result fails its check.
ZB Cloud · runs the approved build
Our cloud runs on every cloud. Each workload goes to the big global clouds, a sovereign cloud, a specialist GPU cloud or a decentralized cloud by price, capacity and location, and moves when one of them changes.
Shared across the stack
What every AI teammate draws on, and writes back to.
Each specialist receives the requirements and product history relevant to its task. Its output is checked before it is added to the wider product.
Technical detailWho does what, and how it is checked.
| AI teammate | Specialists | Does | Checked by |
|---|---|---|---|
| Lead | Plan · Check · Record | The plan, the assignment of every task, the integration of every result, and the record | The read-back with you, and the record against the agreed scope |
| Product | Spec · Research · Roadmap | Requirements, priorities, the roadmap, and the product’s decision history | Review against customer needs and the plan |
| Design | Screens · Flows · Brand | Pages, flows, the interface, and the brand as customers see it | Review against the design brief and the spec |
| Coding | Build · QA · Security | Code and tests, the gates every change must pass, data boundaries and access | The test suite, and a separate review by a model that did not write the code |
| Marketing | Brand & content · Campaigns | The site, the messaging, the content, and the campaigns | Agreed measures of reach and response, in the record |
| Sales | Prospecting · Pipeline | Prospect research, outreach, qualification, and follow-up | Follow-ups and pipeline movement, in the record |
| Operations | Orders · Inventory · Finance · Admin | Orders, inventory, shipping, returns, the books, and the admin a company collects | Activity and cost reconciled against the plan |
| Service | Support · Feedback | Onboarding, answers, resolution, and what customers say | Resolutions and recurring problems, reviewed and fed to the roadmap |
| ZB Cloud · Runtime | Deploy · DevOps | Releases and the path to live; hosting, monitoring and run cost | Agreed service targets and recovery checks; rollback armed on every release |
| Across all | Learning | Outcome capture, business memory, and improvement proposals | Measured results in the record, kept beside the projection |
The roster is open. When a task appears that no specialist covers, a payments integration, a data migration, a store connector, the Lead spawns a specialist for it, gives it the context that task needs, and checks what comes back the same way.
Every result is checked by something that did not produce it: a test suite, a review model, a policy check or a service objective. Work is added to the product only after it passes.
An AI team that works from the same plan.
The Lead breaks your goal into tasks, assigns the work, checks the results and records what happened.
Product keeps the priorities and the business history, and uses customer, sales and support feedback to propose what should change next.
The AI team is built for an industry, so it proposes the systems businesses like yours run on. Twenty for a store. The proposal marks the ones it suggested.
Match the model to the task. Check what comes back.
The Model Router chooses a model as each task starts, from the board of twelve, carries the context forward, and re-routes when an output fails its check.
Tasks from the AI teammates pass through the Model Router, which routes each one along a dotted line to one of twelve models: GPT for drafting and planning, DeepSeek for summaries, Qwen for extraction, Fable for specs, Gemini for images, GLM for translation, Astra for hard planning, Llama for cleanup, Sonnet for testing, Opus for review, Mistral for tagging and Grok for research. Each result passes a check gate marked Checked before it reaches a box marked Finished work, done and shipped. One long build fails its check at Sonnet and is moved to Opus, where it passes. Illustrative routing, not a benchmark.
One task, step by step
This scripted example shows the work behind an abandoned-cart email for a store that sells in two currencies. Each step records its model or service and its estimated model cost.
| Step | Model or service | Cost |
|---|---|---|
| Write the task spec | Fable | $0.22 |
| Plan the files | GPT | $0.03 |
| Write the code, four files | DeepSeek | $0.04 |
| Write the tests | Sonnet | $0.05 |
| One test failed. Re-routed to the next model up. | Opus | $0.08 |
| Review the change | Opus | $0.07 |
| Write the email copy and subject lines | GPT | $0.01 |
| Merge and deploy a preview | ZB Cloud | $0.00 |
| Total | $0.52 |
The same action with everything on one model: $1.45. Illustrative model costs at list rates; hosting and the 5% fee are separate.
Routine tasks
The cheapest model that passes the checks.
Complex tasks
The strongest model for the job.
Images
A model that makes images.
Failed checks
Another model takes the task. The retry and its cost go in the record.
Launch it. ZB Cloud keeps it running.
Our cloud runs on every cloud. Releases, monitoring, backups and recovery are part of it. Each job goes where price, capacity and location say it should. Data that must stay inside a border stays there. Hosting is charged at provider cost.
Work from running the store passes through ZB Cloud, which weighs price, capacity and location and routes each job along a dotted line to one of four clouds: Hyperscaler, everywhere at list price; Sovereign cloud, data stays inside the border; Neo Centers, GPU capacity by the hour; Decentralized Clouds, thousands of independent machines at a fraction of the price. Each job runs where it lands. Patient records always go to the Sovereign cloud. The nightly batch first runs on the Hyperscaler; when the price changes, ZB Cloud moves it to the Decentralized Clouds.
Zero DevOps: the system operates itself. That is what the stack is for.
Keep what the work teaches you.
Your business history
Decisions and results stay together, so the next plan starts from what you already know. The Product teammate keeps it.
How ZB Stack gets better
Tested results change how ZB Stack works. A change needs evidence before it becomes the new way.
See the change, the cost, and the result together.
The record shows each change, why it was proposed, who approved it and what it cost. Measured results are added as they come in. Retries and failed checks are in it too.
In the scripted example, seventy-nine tasks failed a check and were re-routed. The record shows each retry and what it cost.