Cass is a personal AI agent — an in-house JARVIS — that operates a real homelab and cloud footprint end to end. It writes and ships its own code, manages infrastructure, remembers everything, places phone calls, drives browsers, and heals its own failures. This page is the honest technical rundown of how it's built and everything it plugs into.
A long-running service wrapped around a frontier model, given a full toolbelt and a persistent brain. Nothing exotic — just a lot of well-wired parts pointed at real work.
FastAPI + uvicorn service running as a managed system daemon inside a container on a self-hosted Proxmox cluster. Blue/green deploys with a draining restart so a live conversation is never cut off mid-turn.PostgreSQL long-term memory store reached over a Model Context Protocol (MCP) server — hot/cold tiers, full-text search, and topic-scoped recall. The agent reads and writes its own memories as it works.Cass isn't a chatbot with a few plugins — it holds credentials to and operates the real systems below. This is the full surface.
Two separate AWS accounts (personal and business), operated directly and kept strictly isolated from one another.
A self-hosted Proxmox cluster of containers and VMs, orchestrated straight from the agent — provisioning, restarts, resizing, health checks.
Reads, writes, reviews and ships code through Git. Application and Lambda deploys are locked to CI/CD — the agent pushes commits and pipelines do the rest.
Places real outbound phone calls (spoken, and two-way with both sides recorded), holds SMS conversations, and pushes to Chester wherever he is.
When there's no API, the agent drives the screen: a headed browser with structured DOM control, plus native macOS and Windows automation with 2FA scraping.
Fetches logins on demand from a mirrored password vault and the OS keychain, and reads one-time 2FA codes so it can sign into things itself — no passwords pasted by hand.
A signal processor reads new messages across every channel every few minutes, extracts to-dos, and decides when something is urgent enough to interrupt.
Resolves the references in everyday life — money, people, dates — against the real sources instead of asking.
"Download that" means the agent queues it to the right media manager and it lands in the library — no torrent sites, no manual files.
Checks the cameras and doorbells on request — over the cloud, so it still works even when the homelab and power are down.
Plugs into the tools a working IT professional lives in — ticketing, time entries, tenant administration.
A sentinel watches every automation across the fleet and asserts its expected result. When something breaks, the agent's default is to fix it — then report.
Operational specifics — hostnames, addresses, account identifiers and credentials — are deliberately left out of this page. Cass holds them; the public write-up doesn't.
Given "fix X," the deliverable is a pushed commit and a green pipeline — it clones the repo, makes the change, and confirms the deploy landed.
Restarts wedged services, resizes disks, rotates keys, provisions containers, and manages DNS and certificates across a homelab and two clouds.
Calls a vet to confirm an appointment, negotiates by text, books a table, checks the front door camera — with recordings and an audit trail.
Persistent memory plus self-scheduled check-backs mean it doesn't drop long-running work or forget what was decided three weeks ago.
A monitor asserts the expected outcome of every automation; when one fails, the agent diagnoses and repairs it before a human is ever paged.
One notification per event, terse answers, and a strong bias to act rather than ask. It's a capable subordinate, not a chatbot.