PrimeForge

One foundation, many products.

Every PrimeForge product is a different surface on the same machinery. This page is the machinery.

A collection of unrelated tools is only a collection. What makes PrimeForge a family is that each product is a thin domain layer over one shared runtime — the same gateway, the same memory, the same permissions, the same audit trail. Adding a product means adding what is genuinely specific to it, not rebuilding everything underneath it.

That is also what makes continuous change affordable. When the foundation is shared, an improvement to it arrives in every product at once — and a product that stops earning its place can be retired without taking anything else with it.

Frontier models are treated as supply, not as strategy. They made intelligence abundant; what stayed scarce is the ability to direct it safely inside a real business, with real consequences attached. Everything below is about that second problem.

  1. 01

    Context

    The business, regulatory and historical context a general model does not carry.

  2. 02

    Control

    Permissions, approval gates and an audit trail around every consequential action.

  3. 03

    Continuity

    Provider failover and local execution, so capability does not vanish with a supplier.

  4. 04

    Delivery

    Products you install and run, rather than an integration project with a start date.

From context to auditable action.

The same five steps run underneath every product. Only the rules and the domain workflow change.

The PrimeForge decision pipelineFive stages in sequence: business context, then rules and risk, then a decision, then an action, then evidence. A human approval gate sits between the decision and the action, so nothing with a consequence attached leaves the system until a person has approved it.Business contextcompany · people · systemsRules & riskdomain · regulatoryDecisionprioritise · explainActiondraft · task · updateEvidenceaudit · escalationHUMANnothing with a consequence proceeds until a person approves it

A worked example

An agent notices a new obligation. It maps that obligation onto one specific business — its size, its sector, what it already has in place. It prepares a recommended action and the evidence behind it. A person approves. Only then does anything execute into a real workflow, and what happened is recorded where an auditor can find it.

Where the gate sits

The approval step is between the decision and the action, not after it. Machines capture, analyse, draft, recommend and rank; a person approves anything that leaves the organisation or creates an obligation. The organisation still carries the consequence, so the organisation still makes the call.

Why it generalises

Nothing in those five steps belongs to a domain. Swap the rules module and the same pipeline serves a different sector, a different market or a different regulatory regime — without rebuilding anything below it.

Four layers, built bottom-up.

Infrastructure first, then intelligence, then execution, with governance wrapped around all of it rather than bolted on top.

  1. L1

    Cognitive infrastructure

    Multi-model gateway · agent runtime · memory · tool calling · local execution

    The runtime everything else is written against. A product asks it for capability and gets it, without knowing or caring which model answered.

    SHARED
  2. L2

    Decision & risk intelligence

    Business context · regulatory rules · risk scoring · recommended action · evidence

    Where a general model becomes useful to one specific organisation: its situation, its obligations, and what follows from both.

    DOMAIN
  3. L3

    Execution agents

    Research · drafting · scheduling · multi-step workflows · system updates

    The layer that does the work rather than describing it — and the layer where a human sits between the recommendation and the consequence.

    DOMAIN
  4. L4

    Governance & resilience

    Permission · audit · escalation · provider failover · graceful degradation

    Wrapped around all of it, not bolted to the top. Every consequential action passes through the same permissions and lands in the same audit trail.

    SHARED

Read it this way. Layers 1 and 4 are shared infrastructure — every product uses the same runtime and the same controls. Layers 2 and 3 are where domain knowledge lives, and that is where a product, a market or a partner becomes specific without forking the platform.

Local by default, with one controlled exit point.

The technical choice that shapes everything else: the system runs on your own machine, and exactly one component is allowed to reach outside it.

Local execution with a single controlled exit pointCapture, application data and agent orchestration all run inside a boundary marked as the user’s own machine, with traffic circulating freely between them. One component, the LocalAgentGateway, sits at the boundary and is the only route out. Through it passes the minimum context a task needs, to whichever of four providers is configured: the user’s own Claude plan, their ChatGPT or Codex plan, a local model, or a direct API call as a fallback.THE USER’S MACHINE · DATA STAYS LOCAL BY DEFAULTCapture & connectorsbrowser extension · filesbusiness systemsApplication & datarecords · conversationsanalysis · local storeAgent & task orchestrationdrafting · schedulingmulti-step local jobsLocalAgentGatewaythe only controlled exit point · routing · quotas · session reuse · logging · fallbackminimum necessary context onlyYour Claude planalready paid forYour ChatGPT / Codex planalready paid forA local modelno network at allDirect APIfallback only

Predictable cost

You run on the AI plan you already pay for. We do not resell metered tokens, so there is no second meter running behind your work.

No provider lock-in

Workflows are written against the gateway, not against a vendor SDK. Swapping a model — or falling back to a local one — is configuration rather than a rewrite.

Nothing routes through us

We never become the custodian of data we have no business holding. When a task genuinely needs a model, it sends the context that task needs and nothing else.

Four constraints the architecture is designed against.

These are engineering constraints rather than positioning. They are the reason everything above looks the way it does.

  1. 01

    Local first

    Data stays on your own machine or infrastructure wherever it practically can. When a model is required, only the context that specific task needs leaves — never a whole document store or business dataset.

  2. 02

    Human-in-the-loop where consequence matters

    Machines capture, analyse, draft, recommend and rank. A person approves anything that leaves the organisation or creates an obligation. Automation earns its reach one approved step at a time.

  3. 03

    Out of the box, not a project

    Install and run. No implementation phase, no configuration consultancy, and no requirement to understand any of the architecture on this page. For a small business, adoption cost is usually the real blocker.

  4. 04

    Software that adapts around you

    Traditional software asks the customer to bend their process to fit fixed features. AI makes it cheaper to shape the software around the actual use case — and to keep reshaping it as that case changes.

This is the part you should not have to think about.

None of it is a prerequisite for using anything. The architecture exists so that the products can be installed and run by people who never read this page.