Every "my AI stack" post you will read this year has the same flaw: it lists tools the author tested, not tools that survived. Those are different lists. Testing selects for novelty; surviving selects for work actually shipped. So here is my filter, stated upfront — a tool makes this page only if losing it tomorrow would force me to change how I work on real projects: this 500-plus-post Laravel platform, client work across my agencies, 8+ years and 1,500+ projects of accumulated habit. Tiered S to C, graveyard included, because the tools I cut tell you more than the tools I kept.
One de-hyping note before the list. My stack got smaller this year, not bigger. That is the actual trend line nobody selling you tools will report.
The Rule That Makes the Stack Work: Harness, Not Religion
The most important thing in my stack is not a tool. It is a file.
Every serious project I run has a CLAUDE.md at its root, the project's operating manual: conventions, gotchas, verify commands, the "do not migrate this legacy structure" warnings. The repo, not the tool, is the source of truth. Any agent I point at the project reads that file and inherits a year of context. When I trialed a second coding agent side by side, it worked on the same repo with the same context file, and the switch cost was minutes.
That is the harness principle: tools plug into the project, never the reverse. Every tool below was chosen, and every tool in the graveyard was cut, by whether it respects that arrangement. The moment a tool wants your project to reorganize around it, it is negotiating for your future switching costs.
S Tier: Would Redesign My Week If They Vanished
Claude Code
The daily driver, and it is not close. Claude Code runs my development, my content operations, and increasingly my maintenance work, as an agent in the terminal and editor, not an autocomplete. The features that justify the tier are the unglamorous ones: background task execution so long jobs run while I work, parallel subagents for bulk work (I have run six-wave production runs building an 854-item prompt library), skills for repeatable procedures, and hooks that gate commits behind my formatter and static analysis.
That last one matters more than any model benchmark: my pre-commit hook runs Pint and PHPStan on staged files whether a human or an agent wrote the change. An AI stack without a verification harness is a liability generator. The linter, the type checker, the test suite — those are AI tools now, because they are what makes agent output trustworthy at speed.
Laravel Herd
Local environment for every PHP project. Zero-config services, native speed on the Mac, and, the part relevant to an AI stack, an environment so predictable that agents never waste turns diagnosing environment drift. Agents amplify whatever your environment is. If it is flaky, you now have a very fast flake-explorer.
GitHub + Actions
Version control, but more specifically: the deploy pipeline. This site ships through a GitHub Actions workflow that deploys on push to main. Agent-written changes and human-written changes go through the same PR flow, the same checks, the same pipeline. Boring, load-bearing, S tier. The gh CLI is the piece agents use constantly — PRs, issues, review comments, all scriptable.
A Tier: Weekly, and Missed When Absent
MCP Servers — the Real Multiplier
If I could hand one acronym to 2024-me, it is MCP. Model Context Protocol servers are how the agent stops being a text generator and starts touching real systems. Three earn permanent slots:
- Laravel Boost, the Laravel-specific MCP: documentation search, database schema, tinker execution, route listing, straight from the agent. It removes the entire class of error where the model guesses at your framework version's API. Setup notes in my Laravel Boost MCP guide.
- Browser automation (Playwright / Chrome DevTools): agents that verify their frontend work by driving a real browser and reading the console, instead of declaring victory from the diff. This closed the "it says it works" gap more than any model upgrade.
- Figma: design context flowing into code sessions; it powered the token-driven redesign of this blog.
My longer list, with the ones that did not make the cut, is in the MCPs I actually keep installed.
A Second Coding Agent, as Instrument
I keep OpenAI's Codex in rotation — not as a Claude Code replacement but as a second opinion with different failure modes. Same repo, same CLAUDE.md, run side by side when a decision is expensive. I documented a full same-repo comparison in Claude Code vs Codex, tested on identical tasks. The harness principle is what makes this free: because the project carries the context, a second agent costs nothing to try. Verdicts differ by task; owning two instruments beats betting on one.
Claude (Chat)
Strategy, writing critique, thinking-out-loud. Kept deliberately separate from the coding agent: chat is where questions stay questions instead of becoming premature file edits. The discipline of which surface you open is itself a productivity tool.
B Tier: Specialists on Retainer
- Algolia (via Laravel Scout): powers this site's search. Predates the AI wave, now quietly essential to it: retrieval over my own 500-post corpus is infrastructure both for readers and for agents working on the content.
- Higgsfield (via MCP): image and video generation wired directly into the agent workflow, which is the detail that matters: media generation as a pipeline step, not a browser tab. My YouTube video workflow with Claude Code and Higgsfield shows the full chain.
- HeyGen (via MCP): avatar video for explainer content, same pipeline logic.
- Ollama: local models for offline experiments and anything that should not leave the machine. Narrow, permanent role; notes in running Claude Code free with Ollama.
The B-tier test is honest narrowness. Each of these does one job, does it from inside the workflow rather than beside it, and has survived at least six months of me forgetting it exists between uses.
C Tier: Watched, Not Trusted Yet
C tier is a holding pattern with an eviction policy: anything here either earns promotion within a quarter or gets deleted, because an ever-growing "someday" tier is just a graveyard with better PR. Currently: one AI-first editor I re-test quarterly (verdict so far: clever demos, and I still finish the real work in my main setup), and a rotating slot for whatever launched this month. The slot count is fixed. That is the entire discipline.
The Graveyard, Which Is the Actual Content
Tools I paid for, used sincerely, and cut:
- AI-first editor forks. Tested repeatedly, cut repeatedly. They want to be the workflow; the harness rule says no. The agent moved into my terminal instead, and that arrangement has outlasted three waves of editor launches.
- Standalone chat for code. Pasting code into a chat window and pasting answers back is dead to me. If the model cannot see the repo, run commands, and verify, it is speculating — politely.
- All-in-one "AI workspace" apps. Each one wanted my notes, tasks, and docs inside it. Migration in, lock-in discovered, migration out. Twice was enough; plain markdown in git won.
- A half-dozen single-feature subscriptions (transcription here, summarization there), absorbed by the main agent as it gained tools. This absorption is the quiet story of the past year: the agent-plus-MCP pattern keeps eating adjacent products.
Nothing in the graveyard was a bad product. Every one failed the same test: it optimized for being impressive in its own window instead of useful inside my project.
The Five-Question Filter
When something new launches (which is hourly now), it gets this gauntlet before it gets my attention:
- Does it solve a pain I have this week? No → bookmark file, move on.
- Does it work inside my harness (repo, terminal, MCP), or does it demand I come to it?
- Can I test it on real work within seven days? Toy tests produce toy verdicts.
- After the test: would I notice losing it? Noticing is the tier threshold.
- If it enters, what does it displace? Stacks that only grow are collections.
That filter is why this page lists roughly a dozen tools instead of forty, and why I expect the specific names here to churn while the shape survives. Tools are weather. The harness is climate.
Build yours from your own work, not my list — and if the version of this you need is a stack designed around a real product or platform with deadlines attached, that design work is literally what I do for clients: see my services.
