I'm a forward-deployed AI engineer. I build agents that do real IT operations work, and the controls that make them safe to hand access to: policy before every action, a record after, a human on anything consequential, and evals that say whether it actually worked.
By day I lead AI and automation delivery at an MSP in San Diego: service-desk tooling, an investigation agent that drafts cited findings for technicians, production MCP servers, and the Rewst workflows behind a case study on 160+ hours a month recovered. I started on that service desk, which is why I build with the people who close the tickets.
Franklin, my agentic home server
At home I run the same ideas with fewer guardrails. Franklin is a Proxmox server (24 cores, 96 GB, Radeon GPUs) operated by an agent of the same name. I talk to it by voice or Discord, and it has a real root account. It's where I find out how much autonomy an agent can carry and which controls actually hold.
- The machine becomes what I ask for. "Game night" puts a Windows VM with GPU passthrough on the display; other scenes bring up a Linux desktop or an AI lab. An experience runtime plans against the live host, runs dry by default, requires confirmation for guarded actions, and writes a receipt for every apply.
- It works on its own, within limits. A persistent executive reconciles its queue every 15 seconds, samples health every minute, and starts bounded background work under a daily budget. Results go to review with evidence. A worker saying "done" doesn't count as a passing check.
- It asks before anything consequential. Proposals reach my Discord DMs as approval requests tied to one exact revision. Purchases, deletions, firewall changes, and VM force-offs need an explicit yes.
- Root goes through a broker. Every request is logged with the caller, operation, target, and a command hash. It's documented as an authority boundary, not a sandbox, because that's what it is.
- It ships what it builds to a self-hosted Openship instance, with a token scoped to its own projects.
- Local voice (Speaches, Kokoro), Qdrant memory, SearXNG search, and Prometheus/Grafana round it out.
The agent itself is private for now. Franklin Broadcast
is the piece that's public: Discord slash commands (/s-start, /s-swap, /s-play) drive a
persistent Chromium worker in a virtual desktop that joins the call and shares a tab. Franklin
orchestrates the stream but never is the streaming identity.
Open source
| Project | What it is |
|---|---|
| dotshot | Mac app: a screenshot, recording, or file goes to your agent's machine over SSH, and the remote path lands on your clipboard. Swift. 1.0 is out. |
| Cadre | A self-hosted control plane for AI coworkers: each agent gets its own computer, CEL policy decides every action before it runs, and the agent never holds the credentials it uses. TypeScript, alpha. |
| Terminal Brain | Native macOS app that gives agents local-first memory over MCP from Obsidian, Apple Notes, and Drafts. Agent writes land in a reviewable inbox. Swift |
| Franklin Broadcast | Discord browser-streaming automation: Playwright worker, noVNC desktop, stream control API. |
| M365 Utilization Report | Multi-tenant Microsoft 365 license, usage, and cost reporting as a signed Rewst workflow. |
| NotMyRouter | Continuous probes that prove whether your ISP or your router is at fault. Python |
| MiniMix | Per-app volume and mute in the menu bar, using Core Audio process taps. Swift |
Also: a 4-GPU AMD ROCm inference lab with a Python benchmark harness (write-up). Talks: FLOW 2026 Community Live · automation adoption webinar.