My MacBook is the server now

I can run a coding agent from my phone now, but none of the code or compute is on the phone. Oh My Pi runs on my MacBook. Tailscale gives the Mac and my phone a private network, and Termius gives me a usable SSH terminal on a very small screen.

Why did I choose this stack? Because it just works. Others have used Orca, or even the remote options that Claude and ChatGPT offer.

The result is simple - I can check an agent, answer a question, or start work while I’m away from my desk. The project stays on my Mac, there’s no cloud copy to keep in sync, and I didn’t open port 22 on my router.

Here’s the setup:

1. Install Tailscale on the Mac and phone, then sign both into the same tailnet.
2. On the Mac, open System Settings → General → Sharing → Remote Login. Turn it on, choose Only these users, and add your account.
3. In Termius, create a host with the Mac’s stable 100.x.y.z Tailscale address or its MagicDNS name, your Mac username, and port 22. Use an SSH key if you have one; a Mac password works too.
4. Install tmux on the Mac so a dropped phone connection doesn’t take the agent with it.

brew install tmux
tmux new -s omp
omp

Detach with Ctrl-b, then d. Later, reconnect through Termius and run:

tmux attach -t omp

The only failure mode is that the Mac (or device) has to stay awake, online, and connected to Tailscale. If it sleeps, your tiny remote command center becomes a black rectangle. You can take care of this with a simple caffeinate command or updating computer settings.

Why I’m using OMP for it

OMP is an open-source fork of Pi, not a Pi extension. It adds subagents, browser control, LSP, persistent sessions, hash-anchored edits, and 40-plus model providers. Basically anything that cursor, opencode, or other alternatives offer. It’s an extremely capable CLI tool that pairs well with CMUX if you want GUI or browser capabilities. Install it with:

curl -fsSL https://omp.sh/install | sh
omp

Run /login inside OMP to attach a supported subscription or provider account. OMP routes models by role, with custom provider definitions in ~/.omp/agent/models.yml. I’m using GPT-5.6 Sol at high or xhigh for planning, advisory, and design; Luna for routine work; and Gemini Flash for small vision tasks. That lineup will change. The remote path back to the Mac won’t.

The model race moved again

OpenAI reports Sol at 72.7% on DeepSWE v1.1 versus Fable 5 at 69.7%, and 88.8% on Terminal-Bench 2.1 versus 83.1%. On the live Code Arena React leaderboard, Sol was at 1638 ±21 and Fable at 1632 ±17 on July 14. Those confidence ranges overlap, so I wouldn’t turn six points into a victory parade.

Anthropic then extended Fable 5 access on paid plans and kept Claude Code’s weekly rate limits 50% higher through July 19. I can’t prove Sol caused that response. However, the timing makes it an awfully reasonable inference.

Claude’s July 12 extension announcement.

Grok 4.5 is another cheap option worth testing: xAI reports 80 tokens per second, $2 per million input tokens, $6 per million output tokens, and a 29% SWE Marathon resolution rate versus 26% for Opus 4.8 and 24% for Fable. It trails both on several other coding evaluations. Useful, not a sweep.

This is why I want the harness and remote workflow to stay boring. The models will trade places. Changing a role assignment is easier than rebuilding the way I work.

Worth Reading

Routing GPT-5.6 through Claude Code — Kai
A clever five-minute experiment, but the author flags a real Codex account-ban risk. Lab setup, not my recommendation for a work account.

A compiler-based MCP for cutting agent token use — Donny
Early and unproven, with eager indexing that can make the session slower. The CPU-for-token trade is still worth watching.

Supabase skills that pull current product docs — Tyler Shukert
A practical way to give an agent current Supabase guidance instead of trusting stale training data.

Have you tried reaching a local coding setup from a phone or tablet? Reply with the device, terminal, and harness you’re using. I’m curious what others are using.

Forwarded this email? Subscribe here.

— Collin