Desktop app on Windows. This is where I build apps, run audits, write content, and orchestrate everything else. Persistent memory means it knows my businesses across sessions. Most of what you see on this site was built inside it — including this site. Internal, non-PHI work only.

Any time client PHI is in play, Claude runs as Opus 4.6 through AWS Bedrock, under a signed BAA — never the consumer app, never the desktop app. Same model, the compliance footprint healthcare work actually requires.

Career That Lasts runs real Claude API calls for career analysis. When I ship AI features in my own products, this is the backend.

Claude drives a real Chrome session for me — editing GitHub repos, changing DNS at my registrar, checking deploys — by operating the browser the way I would. Anything outward-facing, I confirm before it goes through.

The difference between Claude as a chatbot and Claude as an operator. Every connected service on this page reaches Claude through MCP — it reads my real data and drafts real actions, and I approve them before anything goes out.

Every product I've shipped with Claude Code uses this combo. Claude knows it cold, which matters more than your personal preference does.

Push to GitHub, it's live. Career That Lasts and this site both run on it. Deploying on day one changes how you build — everything gets tested against a real URL.

Claude Code commits and pushes via the gh CLI. The commit history doubles as a build log — it's how I can tell you honestly how long things took.

Node 24 is the runtime behind every Next.js build and script Claude Code runs here. It is pinned and kept off the system path so builds stay reproducible.

The AI Jeff iMessage bot is a Python process: launchd keeps it alive, it reads the Messages SQLite database read-only, sends through AppleScript, and reads its on/off Focus signal from iCloud. Claude Code wrote and maintains all of it.

Self-hosted on Railway, on BAA-covered infrastructure. This is the agent backbone that orchestrates client work, and self-hosting it is exactly what keeps the automation inside our compliance boundary instead of on someone else's shared cloud.

When two apps just need data moved between them on a trigger, a Make scenario beats an AI agent. I design the scenarios with Claude, then Make runs them deterministically — on its enterprise BAA, so even the plumbing is covered for PHI.

Where the self-hosted n8n lives, under BAA. The automation layer runs on infrastructure we hold an agreement for, not whatever was easiest to spin up.

Serverless Postgres behind the n8n agents — HIPAA-enabled, BAA signed. The database the automation reads and writes lives inside the boundary.

Claude reads threads and drafts replies through MCP. It does the triage and the first draft; I keep the send button.

Claude checks availability, suggests times, and drafts events through the connector, so scheduling is a sentence instead of a tab-switch.

Claude reads channels and threads and drafts messages through MCP — useful for catching up on a busy channel without scrolling it myself.

Where longer-lived notes and project docs live. Claude reads and writes pages through the connector, so what it learns in a session lands somewhere I will actually find it again.

Connected file storage Claude can search, read, and create files in, so documents move into and out of a session without a manual upload.

The white-labeled CRM where client relationships actually live — forms, booking, pipelines, email and SMS. It holds client PHI, so it runs under a signed BAA.

Tracks, records, and routes inbound calls under a signed BAA, so call data — which absolutely can contain PHI — stays inside the compliance boundary.

Records and transcribes calls under a signed BAA. Through MCP, Claude pulls transcripts and turns them into summaries and action items without me re-listening to anything.

Our own verification-of-benefits tool for treatment-center clients. It is self-owned and it does handle PHI, so it sits fully inside the boundary.

Connected to Claude via MCP. Keyword research, competitor gaps, and site audit data flow straight into content briefs and audit reports instead of living in browser tabs.

Fast research and sourcing with citations. A non-PHI tool — it stays outside the compliance boundary, so client data never goes near it.

Structured web scraping for research and competitive data. Non-PHI work only, outside the boundary.

For chewing through long documents and research and asking questions across them. Non-PHI material only.

Claude can generate and edit designs through the connector. I use it for social assets where the content matters more than bespoke design.

Claude drafts the narrative, Gamma makes it presentable. Client-facing decks in a fraction of the time PowerPoint would take.

Edits marketing video and podcast audio by editing the transcript. Non-PHI content only.

Stores and serves marketing images and video. Non-PHI assets only.

Campaigns on Facebook and Instagram. Meta never receives raw PHI — everything we send, including Conversions API events for lead generation, is normalized and hashed into an irreversible match key first. The platform can match and measure without ever seeing a real identifier.

Search, display, and programmatic for treatment-center clients, on a signed BAA for the ads and analytics products. Conversion measurement runs on the same hashed-data discipline as Meta.

Hosts humbearmedia.com, the agency site. Content, schema, and redirect snippets live here.

Registrar and DNS for my domains. When godofclaude.com needed to point at Vercel, the records changed here.

Cloud hosting, storage, and the Bedrock endpoint that serves Claude for PHI work — under a signed BAA. The compliant backbone the rest of the stack leans on.

CDN, DDoS protection, and DNS, under BAA. Keeps the edge fast and protected.

Stores credentials, keys, and secrets for the whole stack. Holds no PHI by design.

A mesh VPN tying my machines together, so the team can reach the always-on office workstation — and the rest of the hardware fleet — from anywhere, as if we were all on the same LAN. It is how the big AI workstation gets used remotely without exposing anything to the open internet. Non-PHI infrastructure.

The current main machine: an ASUS ROG Astral RTX 5090 BTF (32GB) with a Ryzen 9 9950X3D2 Dual Edition on a white ASUS TUF Gaming B850-BTF WiFi board, 64GB of G.Skill Trident Z5 Neo DDR5-6000 CL30, and 4TB of NVMe (a Samsung 9100 Pro plus a 990 Pro), cooled by an ASUS TUF Gaming LC II 360 ARGB, on a 1200W TUF Gaming Gold supply, in a black Corsair 4500X BTF with clean back-connector routing. It runs local AI and image-to-video, and the team reaches it over Tailscale. It is the same build featured in the AI PC Shop.

The rest of the custom rigs are the builds in the AI PC Shop: an all-white RTX 5080 (Ryzen 7 9800X3D), an RTX 3090 24GB AI workstation (Ryzen 9 7900X on AM5), and an EVGA RTX 3090 FTW3 Ultra Hybrid 24GB (Ryzen 7 5700X). I build these to run and to sell — some stay in rotation here, some go up for sale.

Three Mac minis — two M4, one M1. One of the M4s runs the AI Jeff iMessage bot around the clock.

Seven or eight MacBook Pro and Air, M4 and M5, for working from anywhere.

A small always-on box that tracks social-media and internal KPIs.