The whole thing.
Every agent, every service, every product — and whether each one is actually running, right now.
One orchestrator. Everything routes through it.
Takes the instruction, decides who runs it, checks the work came back right. Deliberately does no grunt work — the whole point is that it delegates.
Eight of them. Each owns a domain and nothing else.
Every line of code. Builds, debugs, ships. Runs through a bridge that wraps a coding CLI, so it can actually touch the filesystem rather than just describe changes.
Lives on WhatsApp. Handles the group chats and the outreach in a voice that is emphatically not corporate.
Revenue, leads, market and product research. The one asked what is actually worth doing before anyone builds it.
Owns the inventory SaaS as a business rather than a codebase — launch, store listing, support triage, the weekly growth digest.
Sits in the warehouse team's group chat, turns what people promise into a ledger, and follows up until it closes.
Front line for the price-monitoring product. Answers first, escalates second.
Keeps the purchase ledger for the businesses: every card spend logged with its receipt and explained to the bookkeeper on demand. Records, never advises.
Walks the office assistant through one task at a time, checks the screenshot before the next one unlocks, and keeps the timesheet honest.
The terminal agent. Migrations, infrastructure, service plumbing — the work that needs full access to the machine and a human watching it happen.
The services that make the agents more than chat windows.
The switchboard. Every message in or out of every agent passes through here and gets routed to whoever owns that conversation.
The bots the humans actually talk to. Separate identities so a message lands with the right agent instead of a shared inbox.
Reads every session transcript afterwards and extracts the mistakes, decisions and lessons. Recurring problems get hotter until they are impossible to ignore.
Vector recall for things that fell out of a context window. Lets an agent go looking for what it used to know instead of hoarding it.
Wraps a full coding CLI behind a standard chat API, so an agent gets real tool use — file edits, shell, the lot — rather than plain text completion.
Runs on the machine itself. Turns text into vectors for recall without sending any of it to anyone.
The private view of all of this — memory, tasks, schedules, agent health — for when a graph is not enough.
Captures every trade as it opens and closes, then asks for the reasoning while it is still honest. Nothing here trades on its own.
Exists because agents used to claim they had set a reminder and then silently never fire it. Now the claim has to be backed by a row in a database.
What comes out the other end.
Barcode-scanning stock control for small operations. Web and iOS, live and taking subscriptions.
The page you are reading, served from the same machine that runs everything above it. That is why the numbers here are real.
Watches prices and says something when they move. Mira handles the people who ask why.
A monitor on the machine probes every service on this page every ten minutes and writes what it finds to a file. When you load this page, the same server that hands you the HTML reads that file and returns a deliberately small summary: a status, a rough uptime, a model name. The raw results never leave the machine — no addresses, no ports, no identifiers — because the endpoint is built from a fixed list of things it is allowed to say, rather than by hiding the things it is not.
So a red node here means something is genuinely broken right now, and I have not got round to fixing it yet. That seemed more interesting than a diagram that is always green.