▶ YouTube Summaries

• https://www.youtube.com/watch?v=9tGrhrVKCrE
• Original title: Why I'm moving to Linux (for real)

Theo, a long-time Apple fan, explains why he has moved most of his real development work off his MacBook and onto a fleet of Linux boxes. The trigger is the rise of long-running coding agents (Codex, Claude Code): running them locally hammers the Mac's CPU, forces him to leave his laptop open while agents work, and runs into macOS-specific slowdowns. He walks through the concrete pain points that pushed him over the edge, the hardware and networking setup he built to replace it (SSH + tmux + Tailscale, T3 Code as a remote GUI, and network KVMs for full remote control), and argues that with agents doing the setup work, running Linux is now easy and far more pleasant for parallel agent workloads. His core thesis: orchestrating where and how agents run across machines is increasingly the developer's real job, and Linux is a much better OS for it than macOS.

Why I'm Switching to Linux

Setup and motivation. Theo has historically done day-to-day dev work on a MacBook, but that has changed. The biggest driver: running coding agents (especially Codex) locally hammers his Macs — even a maxed-out M5 Max ($10k) machine — to the point of being nearly unusable. Combined with the now-familiar problem of leaving a laptop cracked open so agents keep running, and the fact that his work has shifted toward long, end-to-end agent tasks (make the change, open the PR, wait for review, address feedback), doing this on a MacBook became untenable. He is often away at events/meetings or on flaky 5G in an Uber, and having everything tied to his open laptop was intolerable.

Why not just use cloud IDEs. He likes Devin's and Cursor's cloud implementations, but there's a catch: subscription subsidization. Per semi-analysis, the Claude plans give roughly $400 / $2,000 / $8,000 of monthly inference on the $20 / $100 / $200 tiers, and Codex gives ~$700 / $3,500 / $14,000. For individuals building personal projects, using those subsidized subs on your own hardware is far better value than paying for cloud IDE tokens. So he wants to keep his laptop closeable while still exploiting the subsidized plans — which means running the agents on his own always-on machines, not on his Mac and not in cloud IDEs.

The macOS-specific pain points:

• Codex resource usage — idle is fine, but any real work (especially computer-use) spikes CPU across cores; he's seen a Mac Mini pegged at 100% across all cores, over 100°C. A few concurrent tasks max out even an M5 Max.
• syspolicyd — macOS's system policy daemon monitors every new process for security. Codex spins up huge numbers of processes (each sub-agent needs its own computer-use MCP with ~30+ processes per thread), so syspolicyd monitoring drives CPU insane. Sub-agents make it dramatically worse (fans spin up audibly).
• APFS file performance — a benchmark measuring creation/deletion/recall of many small files (git clean, git clone, pnpm install, spinning up work-trees for sub-agents). On a top-of-line M5 Max MacBook these take 30+ seconds; on a cheap Linux box with ext4 they take under 10s. He measured ~2.5s vs 35s (git clean) and ~7.3s vs 35s (install) — Linux roughly 10–30x faster for exactly the day-to-day work agents now generate constantly.

The Linux fleet workflow. He has multiple machines (Mac Minis, a framework desktop "BB1", an HP Zbook with 32 threads, a ComfyUI rig) all on Tailscale. He accesses them via SSH and via T3 Code (his own tool). Nice touches: SSH auto-attaches to tmux so sessions persist across disconnects. One machine (his MacBook) is the "brains" — it has SSH keys to all others, plus a "fleet" skill and a computers.md file describing every machine, so his agent can SSH in and reconfigure the whole fleet (demo: fixing shell prompts across all Linux boxes to match his Mac's pure setup). Catches with pure SSH: you can't paste images/screenshots over SSH (a real problem for him), and off-home-network addressing needs handling.

Things he gave up and stopped caring about: 128GB unified CPU/GPU memory for local models — he's decided he just uses cloud/open-weight models instead. macOS computer-use (very good on Codex/Mac, e.g. works with display off) — on Linux it's less stable, so instead he gives a Mac Mini a single computer-use task and does the actual work on Linux boxes, with Linux boxes able to SSH into the Mac Mini to trigger computer-use there.

Network KVMs — the key trick. He loves GL.iNet's Comet Pro network KVM (prefers it over the JetKVM for performance). It plugs into a machine's HDMI + USB-C and gives full remote control regardless of installed software — reboot, control the bootloader, mount an ISO and flash a different OS, all remotely. Real story: he used it to recover a broken/non-bootable Linux install remotely by having Codex force a boot in GRUB and repair partitions, unattended. He pairs it with GL.iNet's "Fingerbot" — a wireless robotic finger you stick next to the power button for remote hard reboots. He views the KVM as an emergency interface plus a way to give agents full-computer control, not for daily browsing. He is unaffiliated but effusive about GL.iNet.

Daily driver: T3 Code. He runs T3 Code (e.g. `npx t3@nightly serve`) and connects over Tailscale or LAN to control a machine through a proper working GUI — terminal included, image/screenshot pasting works again, work-trees are effectively free on Linux. He can add remote environments/machines, clone repos onto remote boxes, forward ports to view web UIs, and keep everything running with his laptop closed. He barely uses T3 Code locally anymore (local history 23 days old). This let him shift to much longer autonomous jobs: "investigate the codebase, make the change, open the PR, wait for the review bots, respond to comments, maybe ask Claude for its opinion, and don't bother me until you're done." Another example: audit a legacy project with sub-agents and produce an HTML plan (he built a small microservice to host those HTML plans).

Hardware and cost notes. You don't need a beefy machine — Linux is just better for parallel agent work; a 32-thread box barely breaks a sweat under loads that melt his Mac. He recommends the GMK Tech GMKtec K8 Plus (~$400 barebones, ~$740 on sale with 32GB RAM + 1TB SSD), unaffiliated. Caution: he'd avoid running personal Claude Code / Codex subs on rented cloud boxes — Anthropic doesn't want Claude Code subs run anywhere but personal hardware (ban risk), so keep it on your own local network. Rolling your own "cloud" from an old laptop/desktop or a small NUC gives better utilization, performance, and control.

On the official cloud agent products. He's dismissive of Codex Cloud and Claude Code cloud offerings — hard to set up environments, missing features, poor access to written code, can't swap models, too locked-in; he says they've been de-emphasizing/hiding them in favor of the built-in remote-control options. Cursor and Devin got much closer because they're focused on it, but he still prefers owning the workflow.

Closing thesis. Orchestrating which machine runs which workload — and how to reach it later — is more and more the developer's actual job as we build agents and environments. It's genuinely fun. Setting up Linux is no longer painful because you can just tell Codex/Claude Code to configure the machine for you. He's happier, building more, and finally approaching his usage limits. Teaser: a T3 Code mobile app is coming to control your machines/agents fully remotely (the project is open source). The shared enemy: holding your laptop open while things run (and no, amphetamine is not the solution). He never thought he'd be a Linux dev, but highly recommends the setup.