First, the confession: the April edition of this report promised monthly updates, and then the summer happened. Five months of silence — but not five months of standstill. The stack changed more between April and September than in the year before it, so consider this the catch-up edition. Same rules as always: if I used it, it's in; if I dropped it, I'll say why; nobody pays to be here.
The Big Change: The Publishing Pipeline
The headline move of the summer wasn't a new AI tool — it was ripping out the plumbing under my sites. Both of my content properties came off traditional shared hosting and onto a GitHub → Cloudflare Pages pipeline, with the sites rebuilt in Astro. A commit to the repo is now a deploy: the article you're reading went from draft to live URL in about two minutes, automatically.
The other half of that change is what writes the commits. Claude went from "my primary assistant" in April to something closer to an operations teammate: the site migration, the DNS cutover, the QA passes, and most of this month's articles were built through Claude driving the pipeline end to end while I reviewed and approved. That's the "assign and review" workflow I keep writing about — I'm not describing a demo, I'm describing how this page got here.
The Core Stack
Claude (Anthropic) — Core
Still handling the majority of my daily AI work — writing, coding, analysis — but the shape changed: less chat, more delegation. Claude Code and the agent workflows now publish to my sites directly through the GitHub pipeline. The long context window remains the reason I haven't switched.
n8n (Self-Hosted) — Core
Still the backbone. Tens of thousands of executions a month on the Mac mini across lead gen, content distribution, monitoring, and agent orchestration. Nothing on the market has given me a reason to touch it.
OpenClaw / Molty — Core
My agent framework, still running with the execute-first configuration and the model fallback chain, still with the Telegram emergency reset that has saved me from more than one runaway loop. After this summer's industry-wide agent-containment drama, that reset trigger feels less like a convenience and more like policy.
ComfyUI + FFmpeg — Core
The Night Caps video pipeline hums along: SDXL frames out of ComfyUI, Ken Burns and 3-layer audio out of FFmpeg, 8-hour ambient videos out the other end. Free, local, irreplaceable.
Ollama — Core
Local inference for the small stuff — classification, extraction, template filling. Keeps API costs down and data on my desk.
Mac mini M4 (32GB) — Core
Still the best tech purchase in years; still running the whole self-hosted stack 24/7 at ~15 watts. (Whether it survives the RTX Spark era is a live question — see On Trial.)
Beehiiv — Promoted: Trial → Core
In April this was "on trial." It graduated. The newsletter is now wired into every page of the site, the editor and analytics have held up, and the free tier is still carrying the load. The 1,000-subscriber upgrade tripwire from April stands.
GA4, Apollo + Instantly, Telegram Bots — Core
The supporting cast is unchanged: GA4 with named click events for tracking what actually converts, Apollo + Instantly powering the outbound engine, Telegram as the alert and control layer for everything above.
What I Dropped
Hostinger (web hosting) — Dropped
The big one. Both sites are fully off shared hosting — DNS on Cloudflare, sites on Cloudflare Pages, deploys from GitHub. Hosting cost: $0. Deploy friction: gone. Hostinger stays only as the email provider for my domains, which it does fine. Nothing against them; static sites on a global CDN with git-based deploys is simply a different league.
Cursor — Trial ended
April's trial question was "is this worth $20/month over Claude Code?" and the honest answer, for me, was no. The editor is genuinely excellent — my coding agents roundup still recommends it for people who want the visual-diff seat — but my own work migrated to the delegate-and-review pattern, and Claude Code covers that inside a subscription I already pay for. No hard feelings; wrong shape for my workflow.
On Trial This Month
GPT-6 Astra — Trial (waiting on access)
OpenAI's new flagship landed September 3rd with computer-use numbers that, if real, matter directly to how I work. It's rolling out in waves and hasn't reached my tier yet. The moment it does, it gets the same test everything gets: my actual tasks, against my actual current setup, with the $10/$50 per-million-token price doing the talking.
NVIDIA RTX Spark — Watching (October)
The N1X arrives in October: 128GB unified memory, CUDA, mainstream form factor. If independent reviews back the claims, one of these likely takes over part of the Mac mini's job and becomes next report's headline. Wallet on standby, preorder finger restrained.
The Monthly Numbers
| Line Item | April | September |
|---|---|---|
| Claude Pro | $20 | $20 |
| Apollo.io | $49 | $49 |
| Instantly | $30 | $30 |
| Cursor | $20 (trial) | $0 — dropped |
| Web hosting | ~$10 | $0 — Cloudflare Pages |
| Everything else | $0 | $0 (self-hosted/free) |
| Total | ~$129 | ~$99 |
The trend from April holds: more capability, less spend. The stack got meaningfully more powerful this summer — an automated publishing pipeline is a capability upgrade, not a cost line — and the bill went down. Self-hosting plus free-tier infrastructure remains the cheat code.
What's Next
Three things on the October watchlist: Astra access — the head-to-head test against my current setup the moment it reaches my tier. RTX Spark N1X reviews — memory bandwidth numbers decide whether the Mac mini gets a sibling or a successor. And getting this report back on its monthly cadence — the whole point of a transparency report is that it actually shows up.
That's the stack for September. If you're building your own and want to compare notes, the newsletter goes deeper on individual tool decisions every week.