Skip to main content
News
Platform4 min read

Projects, Roles and Where It Runs

A floor plan that mirrors the business, and a platform that runs where the data has to stay.

Projects, Roles and Where It Runs

In From Cocoa to Chocolate, the gateways and the Orchestrator decide what agents can think with, act on and do. This piece covers the two things around all of them: how the platform is laid out to match the business, and where it runs.

Why a floor plan matters

AI arrives in most businesses team by team, one project at a time. Each project brings its own prompts, models, tools and way of keeping things safe, and the governance is reinvented every time. As adoption grows, so does the sprawl, until nobody can say which models are in use, what they cost, or what each agent can touch.

The answer is not to stop teams building. It is to give every team its own room, with its own rules, inside one building that the business controls.

One floor plan for the whole business

A factory is laid out in rooms, each with its own staff, its own equipment and its own rules about who may enter. Chocolate Factory is laid out the same way, mirroring how a business is actually organised.

  1. 1

    Organisation

    At the top, with its members.

  2. 2

    Projects

    Within it, one for each team or line of work.

  3. 3

    Sub-projects

    Within those, nesting up to three levels deep. A sub-project inherits its parent’s access.

Because access is inherited, the rules set once at the top flow down without being rebuilt at every level. A new sub-project starts with the rules it should have, not with none.

Access, limits and data, set per project

Access per project

Someone can build agents in one project and only use them in another, and that can change at any time.

Limits per project

Each project, and each app key within it, gets its own spend limits, rate limits and allowed models.

Governed by project

Each project’s models, MCP servers, agents, API keys and data stay inside it.

Access is granted project by project, and it moves as people move between teams. One team’s experiment cannot consume another team’s budget. What one team builds is not visible to, or reachable by, another unless the business decides it should be.

The project is where the rest of the platform’s rules attach. The AI Gateway’s budgets and allowed models, and the MCP Gateway’s access to tools, are set per project. Within a project, each app key carries its own limits too, so no single app can exhaust the budget for everyone else.

Because the rules attach to projects, the three questions that keep AI in pilots can be answered one team at a time: which models this project uses, what they cost, and what its agents can touch.

Roles scoped from day one

Together with organisation roles, member permissions and access policies, projects are how roles are scoped from day one. Access is not a fixed list handed out once. It is set per project, inherited where it should be, and adjusted as the business changes.

Building and using are separate rights. The people who shape an agent in one project can be the people who simply use agents in another, so each person has the access their work needs in each place, and no more.

This is what lets a business say yes to AI team by team, without giving every team the keys to everything.

Wherever the data has to stay

For many businesses the first question about AI is not what it can do but where it runs. Regulated industries, public bodies and anyone holding sensitive data have rules about where that data may go. A control plane that only runs in someone else’s cloud does not help them.

Chocolate Factory runs where the business’s data has to stay.

Where it runs

  • In the business’s own virtual private cloud
  • On-premises
  • Air-gapped, with no internet connection
  • In a hybrid of these
  • In any public cloud

The agents, the gateways and the traces live inside the boundary the business has chosen, alongside the environment split and data residency it needs. The record of what every agent did stays where the data does.

That matters because a control plane sees a great deal. Every request, every tool call and every trace passes through it. If the control plane itself sat outside the boundary, governing AI would mean sending the business’s activity somewhere it is not allowed to go.

Models served close to home

The same choice extends to models. Because the AI Gateway can route to any host, a business can prefer models served in-country, by a local AI factory or on its own servers, and fall back to other hosts only when it chooses to.

Data centres and AI factories that want to serve those businesses can list their hosted models in the catalog, set their own prices, and see usage and revenue per model. They gain a place in every customer’s catalog, reaching businesses that need models served close to home.

Sovereignty stops being a reason to stay out of AI and becomes a routing rule.

What it changes

For the CIO and IT, projects and roles decide who can do what, so business teams can build without IT hand-assembling a new platform, and new governance, for each one. For risk, security and compliance, access is scoped by project, and the platform runs where the data has to stay. For finance, every project carries its own budgets and limits.

The floor plan is drawn once, and every room keeps its own rules. See how projects are laid out on the home page.

The short version

  • Projects mirror how the business is organised, and sub-projects inherit their parent’s access.
  • Access, limits, allowed models and data are set per project.
  • It runs in the business’s own virtual private cloud, on-premises, air-gapped, hybrid or in any public cloud.
  • Because the AI Gateway can route to any host, sovereignty becomes a routing rule.