How I actually work with AI right now: pick a model, a harness to run it in, and an interface to drive the harness from. The shift that matters is that AI coding is no longer just chatting with a model — it is a loop: give context, let an agent inspect the repo, make edits, run tests and checks, review the output, iterate. A snapshot, updated when it changes.
Models · Harnesses · Interfaces · Evaluating · Subscriptions · Infrastructure · Tools•Updated
The raw horsepower. Five models, pinned ⌘1–⌘5 in T3 Code.
anthropic · model · default
My default. The model behind most of my Claude Code sessions — the one I hand the large, context-heavy, repo-shaped work to.
+ Strongest at planning and long sessions
+ Careful on niche problems
anthropic · model
The everyday Anthropic model when I do not need Fable — fast, precise, and still the one I trust for tight, deterministic output.
+ Quick and reliable for routine edits
+ Same harness, lower cost
openai · model · second opinion
The heavier of the two GPT-5.6 variants, and my go-to second read through Codex — a different angle when Claude and I keep circling the same answer.
+ Useful contrast to the Anthropic models
+ Good at reading a repo cold
openai · model · codex
The lighter GPT-5.6 variant. What Codex runs for focused edits and the scheduled tasks I hand off to the Codex desktop app.
+ Fast into implementation
+ Cheap enough to leave running on a schedule
moonshot · model · opencode
Moonshot’s open-weight model, run through OpenCode (opencode-go) inside T3 Code. The one non-Anthropic, non-OpenAI model in the rotation — kept around to see where it holds up.
+ Open weights
+ A genuinely different read from the big two
- Still deciding what kind of task it earns
The agent loop itself — the thing that takes a model, gives it tools, and lets it inspect a repo, edit, and run checks. Claude Code and Codex each run their own vendor’s models; OpenCode runs anything. The interface I drive a harness from is a separate choice.
anthropic · cli · coding agent
My default harness for high-context, repo-shaped work — planning, larger edits, and longer sessions where keeping the whole repo in view matters. Runs the Claude models; I drive it from T3 Code or Warp.
+ Strong at planning and repo-shaped work
+ Holds context across longer sessions
- I still pair it with Codex when I want a second read
openai · cli · coding agent
My implementation-and-inspection agent: fast at reading the repo, making focused edits, and running the test/check loop. I drive it through T3 Code or Warp, often as a second read against Claude Code.
+ Quick into implementation and repo inspection
+ Clean inside T3 Code and Warp
- I reach for Claude Code first on the largest, most context-heavy tasks
open source · cli · coding agent
The open-source harness, and the only one of the three that is not tied to a single vendor. I run it through T3 Code on the Go plan, mostly to put Kimi K3 into the same loop as the Claude and GPT models.
+ Vendor-neutral — any model in the same loop
+ Same repo workflow as the other two
- Still deciding how much work it earns beyond Kimi
The surfaces I sit in front of. An interface hosts a harness; it does not replace one.
interface · multi-harness · daily driver
Where I spend most of my time now. Not a harness — it sits on top of Claude Code, Codex, and OpenCode and lets me run any of them from one window, with five models pinned to ⌘1–⌘5 so switching is a keystroke instead of a context rebuild.
+ One interface over three harnesses
+ Switch models and harnesses without leaving the flow
- Warp still wins when I just want a terminal
terminal · interface · daily driver
The terminal, and the second surface after T3 Code. Anything that is not an agent session — git, deploys, the odd CLI run — happens here, and it is still where I drop to when I want the raw CLIs. — my referral
+ Keeps a real terminal in the loop
+ Works naturally with both major CLIs
- Still learning where Oz fits alongside it
anthropic · desktop app
Mostly for chat — conversation and thinking out loud rather than repo edits, which stay in Claude Code. In the rotation, but a smaller share of my week than it used to be.
+ Good for conversation and thinking out loud
- Not a replacement for Claude Code on repo work
openai · desktop app
Where my scheduled tasks live. Recurring jobs run here on a timer while the interactive Codex work happens in T3 Code.
+ Scheduled tasks without a server or cron to babysit
- Little reason to open it outside the schedule
Tools I am testing against my real workflow before promoting them to daily-driver status.
warp · cloud agents
Warp's cloud-agent orchestration. Same family as my daily terminal, so I am testing where cloud agents fit next to Hermes and the local CLIs.
+ Fits the Warp-centered direction of the stack
- Still forming an opinion on when to reach for it
nous research · orchestration · command center
Nous Research’s agent harness, self-hosted on the VPS. It can coordinate Claude Code and Codex as workers and take work over a Telegram gateway, all over Tailscale rather than the public internet. For me it is the remote and async lane — a job I can kick off from my phone — while most real work still happens locally on the Mac.
+ Remote control from my phone via Telegram
+ Private by default — no public exposure
- Still deciding what actually belongs in a remote lane versus the local loop
The paid plans that make the stack work — a different kind of thing from the agents and models they unlock.
anthropic · $100 plan
The plan behind my Claude Code, T3 Code, and Claude desktop usage — enough headroom for the heavier, repo-shaped sessions.
+ Room for longer Claude Code runs
- Still deciding how much belongs in desktop vs. CLI
openai · $100 plan
The plan behind my Codex work and GPT-5.6 access. I am still confirming the cleanest public-facing name for the exact Codex tier.
+ Makes Codex a real daily agent instead of an occasional backup
- Naming around the Codex tiers is easy to make too messy
opencode · Go plan
The plan behind OpenCode and my Kimi K3 access. Cheap enough to keep a third harness in the rotation without thinking about it.
+ Puts an open-weight model in the same loop as the big two
- Only earns its slot if OpenCode does
warp · Build plan
The terminal plan behind my second surface.
+ Worth it for a terminal I open every day
- Oz is still in the evaluating bucket for me
Where the work and the sites actually live — the local box, the always-on box, hosting, and DNS.
apple silicon · local environment · hardware
Where the work actually happens. Not an AI tool, but the machine the harnesses edit on and the checks run against, and the default over the VPS for nearly everything. Still plenty of headroom for the whole stack of CLIs, editors, and a browser at once.
+ Handles the full local toolchain comfortably
+ Quiet, cool, and always ready
- Heavier model and container work still belongs on the VPS
hostinger · KVM 4 · Ubuntu 24.04
My always-on server — a Hostinger KVM 4 box on Ubuntu 24.04. It complements Cloudflare rather than replacing it: Workers stay the default for public web apps, while this runs Docker workloads, private tools, background jobs, and Hermes. Access is private-first over Tailscale, with only SSH and HTTP/S exposed publicly. — my referral
+ Always-on home for self-hosted services and agents
+ Private-first access over Tailscale keeps the surface small
+ Cheap, predictable place to run long-lived processes Workers are not meant for
- I own the patching, backups, and uptime myself
hosting · next.js · opennext · edge
Where my public sites live, including this one and morningportion.com. Next.js apps ship through the OpenNext adapter, so the same App Router codebase runs at the edge instead of on a Node server. Static things skip the framework entirely and deploy as plain assets. Moving here from Vercel put hosting, DNS, storage, and access control on one bill and one dashboard.
+ One platform for hosting, DNS, KV, R2, and access control
+ OpenNext runs a normal Next.js codebase at the edge with no rewrite
+ Static sites deploy as assets with no framework overhead
- The Workers runtime is not Node, so some npm packages need a compatibility flag or a swap
zero-trust · auth · sso
How my private apps stay private without building auth. A one-time-code login sits in front of the whole origin, so tools like my workout logger and a few family apps are reachable from any device but open to nobody else. It is the reason those projects can be genuinely useful to me and still never expose a login form to the internet.
+ Real access control on a personal project without writing an auth system
+ One-time email codes mean no password to manage
+ Applies to the origin, so nothing leaks if a route is forgotten
- Anything behind it cannot be shown off without a separate sanitized build
dns · domains · registrar
My DNS and domain registrar — the layer that points every domain at Workers, the VPS, and everything else. At-cost domain pricing and fast DNS make it the quiet foundation under the whole stack.
+ At-cost domain registration with no markup
+ Fast, reliable DNS for every property
+ Single place to manage all my domains
Companion utilities and helper surfaces that support my workflow.
voice · dictation · macOS
Push-to-talk dictation surface. Hold the hotkey, talk, release — audio is transcribed then processed by an LLM mode of my choice. — superwhisper.com
+ Push-to-talk hotkey makes voice feel like a real input mode
+ Choose transcription and processing models independently
+ Modes for verbose dictation, summarized meeting notes, and clean prose
- Mode customization takes a beat to internalize when you build out more than a couple
notes · vault · editor
My personal knowledge vault and note-taking surface; central to workflows and long-form drafting.
+ Local-first, extensible with plugins
+ Great for Zettelkasten-style linking
vpn · network
Zero-config mesh VPN that keeps my machines and VPS reachable securely.
+ Simple secure networking across devices
+ Works well for remote agent access
browser · privacy
Lightweight browser I use for isolated browsing and testing; useful for quick checks.
+ Fast, minimal surface for quick browsing and testing
menu bar · codex
Menu-bar runner for quick Codex prompts and snippets without opening the full app.
+ Convenient for one-off prompts and small snippets