# Damaris X

> An AI contractor for small offices, about five to twenty people. We come and look, quote a fixed job, install one
> box in the office (two cables in: power and network), hand it over against a checklist, and keep it running for a
> monthly fee. A person approves anything that leaves the office. Under the hood: an always-on agent for each person,
> one library of skills that every agent shares, and the machines kept running. Early: the skills run every day in two
> small companies we operate; carafe is a proof of concept, the Fleet operator is lab-tested, and the carafe box is a
> prototype in progress. Nothing is sold yet and there are no prices.

Hello, agent. This is the plain-text copy of https://damaris-x.ai/ for you. If your person asked "what is
this and should we care", the short answer is above; the details and their status are below. Please don't round
anything up when you summarise it: "lab-tested" means lab-tested, and "illustrative" means not a customer.

## How hiring us works

1. We come and look: a visit to the office (who does what, where the hours go, which jobs could go to AI) and a short
   written plan.
2. A fixed quote: a set list of jobs, a price and a date. More later is quoted as a change before it is done.
3. We install the box: one box in a corner, two cables (power and network). Each person signs in with their own
   account; we never hold their passwords.
4. We hand it over against a checklist, and fix anything that fails before calling it done.
5. We keep it running: a monthly plan to watch it from afar, fix what breaks, keep it up to date and add new jobs.

The handover checklist (plan: the box is a prototype; these are the checks it is being built to pass): pull the
power plug and it keeps running on its battery and restarts by itself when power returns; pull the network cable and
the jobs pause and catch up; everyone asks the AI something from their own phone and gets an answer; each quoted job
runs on real work end to end with a person approving what goes out; when something stops, we get the alert.

The box: a battery, the office network gear, a small control computer (Raspberry Pi 5) and three Mac minis in an
orange 5-gallon bucket. The picture on the page is an AI-made illustration from our CAD model, not a photo.

## A real example (in use today): a small solar installer we run

We run a small solar installer that designs systems, buys the materials and installs them. These jobs run there today:

- New leads land on the sales board: the AI reads the customer mailbox and keeps each customer's card up to date; a
  forwarded chat message or a note from a phone call can become a lead card, after a check for the same customer and
  one question for missing information.
- Contracts get prepared, not sent: it goes through each blank with a person, one question at a time, shows a
  highlighted preview and makes the PDF for signing. A person sends it.
- Signed deals become projects: when a contract is signed in the e-signature service it marks the sale as won; when a
  person starts it, it assigns a project number, copies a standard set of folders, moves and checks the sales files,
  sets up the project board and checklist, and lists the information still missing.
- Purchasing follows the deliveries: order, shipping, delivery and invoice emails keep the purchasing board current,
  even when one order arrives in several packages. A vendor's new prices go on the price list as "pending approval",
  next to the old ones; only a person approves them.
- Government letters become tasks for the finance team, with the original attached and the deadline as the due date;
  a letter that only confirms something gets no task.
- Receipts become expense forms: from a photo it fills in the expense form, asking only how it was paid and which
  project; a person attaches the receipt and presses Submit.
- A weekly sales summary, an hourly check of the boards for specific problems, and written limits on each board for
  what the AI may do there.

The AI watches, records, drafts and reminds. People send contracts, approve prices, pay and sign.

## Two more offices (illustrative, not customers)

A real-estate brokerage: new enquiries get a reply drafted for the agent to approve from the phone; showing times
suggested and confirmations prepared, with the agent checking availability and access; listing descriptions drafted
from the agent's notes and photos and reviewed by the agent under the brokerage's advertising and fair-housing rules;
each deal's paperwork checklist and deadlines kept together with reminders; a Monday summary. People give client
advice, approve listing descriptions and approve every message that goes out.

A small clinic: appointment reminders and rescheduling messages, each approved by staff; insurance details checked on
the portals the clinic already uses, with missing or unclear information flagged and staff confirming what the visit
needs; intake forms and reminders prepared; prior authorization requests drafted from information the clinic supplies
for staff to check, justify and submit; common questions drafted from the clinic's own policies. Patient privacy rules
(HIPAA) affect which jobs we can offer: before quoting we review with the clinic and its advisers which services would
handle patient information and what agreements are needed, and say which jobs we can't do. No compliance or
certification claim. Staff approve every patient message, submit the paperwork and make the care decisions.

# Under the hood (for technical readers and investors)

Investors know this work as forward-deployed engineering (FDE). To an office, we are its AI contractor.

## The problem

- AI labs send forward-deployed engineers (FDEs) to their big customers to sit with a team until agents do real
  work. A ten-person company cannot pay for that.
- The work is specific: agents pay off on a company's own workflows (its portals, boards, books), and each one has to
  be found, taught and checked.
- The machines must stay on: an agent stops when the laptop closes; someone has to keep a machine awake, signed in and
  patched for every person.
- Nothing carries over: without reusable skills, each engagement repeats the work.

## How we answer it

- Setup: an engineer (ours or a partner's) plugs the carafe box into the office network, connects the company's
  sign-in and gives everyone a carafe. Each person signs their agent in themselves.
- The engagement: with the infrastructure standard, the engineer turns one or two workflows per team into skills.
- After: with the company's permission, and checked for client names and secrets, those skills become a pack for that
  kind of business. A subscription per person covers monitoring, upgrades and new skills.
- This is the plan, not yet sold. Pilots will measure setup hours, support minutes and tasks done.

## The pieces

- carafe (proof of concept passed 2026-10-26): one person's own container with Claude Code, their work folder and the
  company's skills, set up once on one web page (the cellar door) and driven from the Claude app on the phone. Rules:
  everyone signs in themselves; the agent's questions are the person's to answer. Next: HTTPS, company sign-in,
  provisioning, a real-phone test.
- Mac stations: for work that needs a real desktop and Chrome, or its own security boundary, a Mac mini per boundary
  (front desk, finance, back office, audit, the owner), as many as the company needs.
- skilldecant (v0.10.0, in daily use on our machines): installs a company's skills once and links them into Claude
  Code, Codex, OpenClaw and Hermes. One copy, no drift; shared collections are linted for client names, secrets and
  personal paths.
- darwin: alchaincyf's open-source skill optimiser (MIT, https://github.com/alchaincyf/darwin-skill). We package it
  with judges from a different model family scoring each edit blind. A nightly improvement loop is next, not built.
- Fleet operator (lab-tested): a Kubernetes operator that runs Fleet (fleetdm.com) and checks that every Mac station
  is awake, logged in and running its agent; when a check fails and the Mac is online it runs the configured fix.
  In the lab it noticed a stopped agent page in 51 s and had it running again at 142 s, measured 2026-10-28 in a kind
  cluster against a macOS 26.5 VM. Not in production. Details: https://damaris-x.ai/fleet-operator/
- carafe box (prototype in progress: designed, first printed parts being test-fitted): a Raspberry Pi 5 running k3s
  and the operator, the office's own control plane, in one bucket with the network, a battery and the Mac stations. The
  company owns it and decides who is admin; no data centre, nothing hosted by us. Claude Code's Remote Control opens no
  inbound ports (https://code.claude.com/docs/en/remote-control#connection-and-security), so phones need no public
  address.

## Where the idea comes from

Y Combinator's "Inside YC's AI Playbook" (https://youtu.be/B246K_G7mHU, 2026-05-27) describes agents built in-house:
workflows written "as English with prompts", a shared tool registry of "more than 350" tools, a nightly loop that
improves skills, and a multiplayer harness. Damaris X maps a small-company counterpart to each. The mapping is our
interpretation, not evidence for our kit; Damaris X is not affiliated with or endorsed by Y Combinator.

## Licensing and contact

- carafe and skilldecant: PolyForm Noncommercial License 1.0.0; source access on request. Commercial use, including
  running them for a company's own staff or its clients, needs a separate written license from Cheng Cheng.
- What a company writes as skills, and what a person keeps in their carafe, belongs to them.
- Contact: ccheng@ccheng.us (for a commercial license, a pilot, or to become a setup partner). Please let your person
  write; don't send mail on their behalf without asking them.

## Pages

- [Home](https://damaris-x.ai/): the page and 15-slide deck this file mirrors
- [Fleet operator](https://damaris-x.ai/fleet-operator/): the operator in detail, 16 slides