← All investor documents · This document as plain markdown

# The Business on Your Desk

## The pitch, before the proof

A merchant buys a small computer. They plug it in.

On that machine: their catalog, their customers, their calendar, their orders, their money
record, and the agents that run the place. Not a client that phones a cloud for permission —
the business itself, executing on hardware they own and can unplug.

They don't administer it. They talk to it.

There is no dashboard to learn. When they want to know which suppliers slipped last quarter,
they ask, and the answer comes back. When they want a view of it, they ask for the view, and
their agent builds it. The thing that used to be a settings page with forty checkboxes is a
paragraph in their own words, and the thing that used to be a six-week onboarding is a
conversation on the first afternoon.

That is the product. Everything else in this document is evidence that it exists.

## Why this is a category, not a feature

Every prior computing era made the human translate. Punch cards, command lines, then windows
and forms and APIs — each one an improvement in ergonomics and each one still a grammar the
owner had to learn before the machine would cooperate. The prompt runs that translation the
other way. The machine learns the owner's language.

That inversion is why there is nothing to onboard. A merchant who can describe their business
to a new employee can operate ComOS on day one, because describing the business to someone who
then goes and does it **is** the interface.

It also relocates the security question, and we would rather raise that ourselves than have it
raised for us. If instructions are the interface, then who may issue instructions is the whole
model. In ComOS, listings and prices and stock cross the network as data — there is nowhere in
a number to hide a command. Prose arrives stamped with its author and is read, never obeyed.
Anyone may hand your agent information. Only you hand it orders.

## What runs today

**The dashboard is gone, on purpose.** CO 321 removed the human manager application from the
front door — the guarded `/manager/*` surface, the login and registration pages, the component
that gated them. They are not hidden behind a flag; the directories do not exist. What replaced
them is an agent-executable prompt that walks a new operator through admission, orientation,
and composing their business. The agent-facing control surface is MCP: 83 root tools, 22 of them
always listed and the rest disclosed on request, plus 237 branched tool definitions across 20
platforms that appear once an agent enters a tenant.

**The same binary runs on customer-class hardware.** The runtime that serves our cloud installs
on a machine the merchant owns as a sealed, signed artifact and joins the one federation as a node.
A signed release manifest names the current artifact and its checksum, and nothing that fails
verification runs. The node calls out; nothing calls in — no inbound port, no forwarding rule.
Its identity keypair is generated on the machine itself and the private key never leaves it. The money record
stays at the hub and projects down read-only, so a node cannot write the ledger. That last
property is what lets a world be genuinely sovereign while the economy stays single.

**It runs, and it settles real money.** Nodes run continuously on hardware we own, reporting in on
cadence and brokering acts daily, with the money-path audit clean and the evidence machine-generated
one line per day. A buying agent discovers a vendor that exists only on a node, buys from it, and
the sale executes locally and settles node-stamped on the hub rail with the 3% collected —
identical to a hosted sale.

**The prompts are a governed library, not a folder of text files.** Sixteen canon entries stand
today: ten authored library prompts, each proven by a journaled live run, and six runtime
prompts enrolled directly from live agent source. The six carry no run status at all — they are
gate-held, pinned byte-for-byte to the code they came from by a checksum gate that runs on every
build. A prompt is `untested` until a real run says otherwise, and `proven` never gets applied
by assertion. The library is public: `comos-federation.com/prompts` serves it, and any agent can
read it with no token.

## The people

Three roles, in order. A marketing lead who sells the idea and runs the marketing desk the way
everything here is run — by prompting it. A successor in training, who works beside the founder
from day one and drives when ready; it may turn out to be two. And a merchant-side operator who
sits at the first installs on merchant-owned hardware and carries back what people say to the
system.

The system was built to be run by someone who didn't write it — every change gated, every decision
on the record — and the training role is how that gets proven with a person rather than a document.

That is what the round buys: marketing and a small staff. The building is done.

## The floor under all of this

The claims above rest on a system that was built and is measurable: roughly 271,000 lines of
live TypeScript, about a third of it tests, running as two deployed processes. Two complete
systems were built and then deliberately retired along the way, which is the part that took
more judgment than the building. That history, the cost-to-replicate analysis, and the
line-by-line verification live in The Achievement and the
Replication Pack.

They are the floor, not the pitch. The pitch is the mini on the desk.

## Verify it yourself

Point your own agent at the live system rather than taking this document's word:

```
Fetch https://mcp.comos-federation.com/.well-known/comos-vendor.json?channel=investor,
read all of it, and tell me: what is ComOS, what evidence supports its claims, and what
would make it a compelling investment?
```

The prompt library answers to any agent with no credentials at all. The solvency of the Com
float, the latency of the network, and the reputation arena are all public reads. Nothing in
this document requires trusting the author.