PRIME OS — Every AI Product Installs an Intercom. We Handed Over the Keys.

Think about how every AI product you've used actually treats its AI. It's an intercom at the front door: you can talk to it, it can talk back — brilliantly, even — but it can't come inside. It describes the room; it cannot cross the threshold. "You should open X and do Y," it says, and then you open X and do Y, because the assistant lives in a sidebar and the building doesn't know it exists.

Now flip the arrangement. Give the AI residency. Let it see which windows are open, remember yesterday, open an application by itself, and run work overnight while you sleep. That's the question PRIME OS was built to answer: what happens when the environment is addressable by the model — when an answer can be an operation instead of a suggestion?

This one is a different kind of case study, because the client was us. PRIME OS is Rocket Logic Global's first-party product — what we build when nobody's writing the spec but our own convictions.

The Product

PRIME OS is a complete operating environment in a browser tab. You sign in, watch a boot sequence, and land on a desktop: icons, widgets, a taskbar, four virtual workspaces, a global command palette, a clipboard manager, notifications — and fifty registered applications opening in draggable, resizable, snappable, tileable windows. Terminal, Files, Calendar, Mail, a spreadsheet, a canvas, a Kanban board; a wallet, a vault, a marketplace; social, chat, and streaming; a GitHub App integration, a schema designer, even GeomC — a small purpose-built language with its own lexer, parser, and compiler, editable inside the OS. At 1024px the entire shell swaps to a mobile launcher over the same applications and the same data — a different shell, never a shrunken desktop.

But the distinguishing feature is the resident. The AI layer isn't a chat panel adjacent to the work — it's a system service with keys to the building.

The Architecture of Residency

Three design decisions make the keys real:

Applications never import each other — they speak through events. A typed publish/subscribe EventBus carries every cross-application message: window requests, file events, wallet transfers, calendar reminders. When every capability is an event, then the AI drives the whole OS through one interface instead of fifty special cases. The intercom-to-keys upgrade is, mechanically, this bus.

Model output is executable. When the assistant's response contains an action tag, the OS parses it, executes it through the EventBus, and re-renders it as a clickable chip inline in the prose. Ask ROKCAT — the resident assistant — how to wire a webhook, and the answer arrives with CloudHooks already open behind it and a chip in the text to jump back. Even bare mentions of application names become navigable chips. The assistant's prose is interface.

The resident remembers and works unattended. Conversations persist with auto-compaction; durable facts live in a categorized memory store; every request carries live desktop context — open windows, active workspace, recent activity. Beyond conversation, BotLab registers bots with three trigger classes — autonomous (event-driven), scheduled (cron-dispatched), and external (API-key authenticated) — and an agent runtime executes multi-step, tool-using tasks with every step written to an audit trail. When a scheduled bot runs overnight, then the morning brings a readable reasoning trace, not a mysteriously changed value.

And underneath, the security spine we put under everything we ship: 38 tables with row-level security enabled on every one — 128 policies — roles isolated in a dedicated table behind a security-definer check, and user-supplied AI keys encrypted with AES-GCM under a PBKDF2-derived key, decrypted only server-side, never once returned to a browser. An operating system that can act on your behalf must be built like it knows what that means.

The Model Layer: One Router, Four Providers, No Favourites Coded In

PRIME OS doesn't have a "Claude feature." It has a single model router that every AI surface calls — assistant, terminal AI shell, agents, bots, social generator, mini-app generator — with Anthropic as one of four first-class providers behind it, adapted natively rather than approximated: system prompts hoisted to the top-level parameter, tool use translated faithfully in both directions, SSE streams normalised so the browser receives an identical token stream regardless of origin, and automatic fallback when a provider errors. Pick your provider in one dropdown, and you've changed the behaviour of the entire operating system.

Today the model layer is bring-your-own-key — your Anthropic key is validated against the live API before acceptance, encrypted, and from that moment every surface in the OS runs on the Claude model you chose. And we'll state the roadmap as plainly as we state everything: BYOK is a real adoption barrier, and the published rollout plan moves through baseline instrumentation, a first-party Claude default path with per-user quotas, Claude-native optimisation (prompt caching on the stable system prefix, native parallel tool use, model-generated memory compaction), agent depth with per-tool permissions and approval gates on irreversible actions, hardening — and finally distribution: a one-click clone path so anyone can stand up their own instance. Measurement-gated at every phase, because Phase 0 exists so we never have to guess.

The Results

A complete, live operating environment at os.rlgix.com: 50 applications, a real window manager with four workspaces, a genuine command interpreter with pipes and an AI shell, and a full mobile shell over the same data
The AI-with-hands thesis shipped as working software: desktop-state awareness, executable Action Chips, persistent memory, and an audited autonomous agent and bot runtime
Provider-agnostic by architecture: one router, four first-class model providers, live-validated BYOK, automatic fallback — the model is a setting, not a dependency
Security as a first-class subsystem: RLS on 100% of public tables (128 policies), definer-scoped roles, AES-GCM key custody with server-only decryption
A published, phased path from BYOK to a first-party default — instrumented before, during, and after

Why It Worked: The Right Step Was the Bus

The Rocket Now principle applies to our own products hardest of all: a single step in the right direction is worth exponentially more than many steps in the wrong one. The wrong steps were the industry's defaults — bolt a chat panel onto each app, wire the assistant into features one integration at a time, forever. The one right step was the EventBus: make every capability in the OS speak one language first, and the AI gains the whole building the moment it learns that language. Fifty applications, one door, one set of keys. When the environment is addressable, then intelligence becomes an operator — and everything after that is just adding rooms.

Still talking to your AI through an intercom? Come see what residency looks like. → os.rlgix.com · RocketNow.com

Editor's note — before publishing: