Two machines. One AI operation.

Chris's Home AI Systems

A Lenovo desktop runs a network of ~30 specialist agents on GPT-5.5, led by Neo. An NVIDIA GB10 runs five local models and its own resident agent, Trinity. An encrypted private link ties the two together — and a stack of watchdogs, loops, and review gates keeps the whole thing honest while it works.

30specialist agents

453skills in play

6AI models

20+live public sites

The mapTwo machines, one encrypted link, and everything that runs on them.

The gatewayLenovo IdeaCentre M68Runs GPT-5.5 through an OAuth subscription — no per-token API bills. Home of the OpenClaw gateway, all ~30 agents, and 20+ public sites behind a single tunnel.

encrypted private link

The muscleNVIDIA GB10Local model workstation with its own resident agent, Trinity, running fully local on Nemotron. Hosts the five models below — no cloud, no API fees — and serves internal services over the private link.

nemotron-3-nano:30b24 GB

Trinity's brain — the GB10's resident agent runs fully local on it

262k context

qwen3-coder-next51 GB

Heavy coding model; also answers visitor questions on the public data academy

largest on disk

llama3.2-vision:90b54 GB

Deep-verify worker — full-detail pass on flagged card scans

deep pass

qwen2.5:32b19 GB

"Griffey" — comp-pricing worker that reasons over live card listings

32k reasoning ctx

llama3.2-vision:11b7.8 GB

Triage worker — fast OCR and visual scan of card fronts and backs

up to 4 concurrent

GPT-5.5Lenovo

The agent network itself runs on GPT-5.5 via OAuth on the other machine

subscription

Agents vs. workers: an agent (like Trinity) carries memory, skills, and a schedule of its own. A worker is a single model with one job — sometimes with a name, like Griffey, who prices cards all day and does nothing else.

How work flows

Text Neo. The system figures out the rest.

The routing rule

One agent for simple jobs. Ebony fans out the big ones.

Defense in depth

Three watchdogs ask three different questions.

Is the process alive?

systemd supervises every site service and restarts it the moment it dies.

instant

Is the right page being served?

A watchdog fetches each site and checks the actual content, not just the status code.

every 2 min

Is anything failing quietly?

Ebony sweeps every site: service state, HTTP response, and log error patterns — then messages Chris only if something is really wrong.

every 2 hrs

Why the overlap existsA review function once crash-looped silently for a week while the process stayed alive. The current system checks process, content, response, and error patterns.

Always running

The system works whether anyone is watching or not.

every 60 secTrinity output watcher

On the GB10, a watcher catches Trinity's exploration output and files it — a checker built for a local model's quirks

every 10 minRothschild trading

Reviews the market and its 25 paper-trading algorithms, replacing losers on its own

overnight + recurringMike/Gabe item swarm

Multiple Mike agents click through website item banks overnight while multiple Gabe agents score, review, and finalize the work in parallel

every 45 minNeo journal

Captures what happened and folds it into working memory

3× dailyCindy check-ins

Runs Chris’s personal operations: inbox, calendar, follow-ups

3× dailyNeo outreach

Morning, midday, and evening runs batch the background work

each workdayTrinity ↔ Neo exchange

The two lead agents discuss a topic in a shared Telegram group Chris monitors. Each opens one per workday, staggered so they never collide

dailyGateway recycle

The agent gateway restarts itself on a schedule — before failure, not after

weeklyOmega strategy

A slower loop that reviews how the whole system is working

weekly + monthlyNeo memory upkeep

Compaction and contradiction sweeps keep long-term memory honest

Sorted fastest to slowest. Every loop runs unattended and recovers on its own if a cycle fails.

The item factory

AI agents run the item banks all night. Other agents review the work.

The important part is not just that an agent can make an item. The system can spawn several Mike agents at once to keep the website banks moving overnight, then spawn several Gabe agents at the same time to score, review, and finalize the output. That is the story: parallel agent labor, not a single chatbot doing one task at a time.

The roster

Every agent, and every skill it carries. Click one.

Agents30

Public skills453

Specialist skills102

Shared utilities29

Shared utility beltCommon tools stay quiet so specialist skills lead.

Checkers and self-healing

The system protects itself before it asks for help.

Self-healing coreRestart, verify, scan, escalate

Memory guards, proactive gateway recycling, fallback provider chains, and escalation rules reduce silent failure without exposing operational controls.

Memory pressure

earlyoom and a low-memory monitor kill runaway processes before the machine hard-freezes — added after it actually happened once

Provider fallback

Market-data sites chain CoinGecko → CoinPaprika → Coinbase, so one dead API never blanks a page

Escalation paths

Ashley hands legal questions to Skip. Neo hands multi-agent work to Ebony. Nobody improvises outside their lane

every 2 minSite watchdog

Fetches every public page and verifies the served content

every 2 hrsAll-sites sweep

Ebony: service state + HTTP response + log error scan

every 2 daysSecurity health

Knight audits the servers and their exposure

weeklyAcademy audit

Ebony re-reads the training site for content drift

continuousPer-site timers

High-traffic sites carry their own extra healthcheck lane

continuousCard queue watch

Cade watches the GB10 pricing queue across the private link

The controls

Four ways in. One person holding the keys.

From a phoneTelegram → Neo or Trinity

Day-to-day driving. Chris texts Neo on the Lenovo or Trinity on the GB10 — each machine's lead agent is one message away, and the routing rule above takes it from there.

Agent to agentThe group chat

Neo and Trinity also share a Telegram group with Chris. The two lead agents discuss ideas with each other — machine to machine — while Chris reads along and steps in when he wants.

From a terminalClaude Code

Anthropic's coding agent, used to build, fix, and deploy directly on both machines.

From a terminalCodex

OpenAI's coding agent — same jobs, second set of eyes. Two vendors, no single point of dependence.

Owner-only authorization gate

Whatever the channel, only Chris can authorize file or system actions. Visitors see the architecture, never the controls.

Ground rules

Four rules keep thirty agents from becoming chaos.

Redundant by design

Every important service is watched by more than one independent mechanism.

Content beats uptime

A live process is not enough — a checker has to see the correct page. A site once crash-looped for a week while its process stayed "alive."

Escalate late

Loops repair and verify first. Chris gets a message only when something needs a human.

Owner-only control

Only Chris can authorize file or system actions. Visitors see the architecture, never the controls.