The stack

What I build the systems on this site with, and why each piece is there. A tool list on its own teaches nothing - the reasoning is the useful part.

Agents and orchestration

Where the work actually gets done.

Claude Code: The orchestration layer. Skills and plugins turn a repeatable playbook into a command, so the work that used to need a hire becomes something I invoke by name.
Claude Agent SDK: For agents that run outside a terminal session - scheduled jobs, long-running loops, anything that has to survive me closing the laptop.
Model Context Protocol: How an agent reaches a real system. One protocol instead of a bespoke integration per tool, which means a customer's own AI can drive my software rather than mine driving theirs.
Markdown on the filesystem: Agent memory. Plain files, in git, greppable and diffable. A database would be faster to query and far worse to reason about when something goes wrong.

The half nobody demos, and the half that decides whether it can ship.

WorkOS: SSO, SAML, OIDC and SCIM without building an identity provider. Enterprise auth is table stakes the moment software touches someone else's data, and it is not where the interesting work is.
Stripe: Payments and Connect. Anything that moves money sits behind an approval gate - an agent can prepare the charge, a human releases it.

Cheap, fast, and boring on purpose.

Cloudflare Workers and Pages: Edge runtime and static hosting on one platform. This site is Pages; the agent-facing services are Workers.
Neon: Postgres with branching. A database branch per environment means a migration can be tried somewhere real before it touches production.
Astro: Static by default, ships almost no JavaScript. For a content site the correct amount of client-side framework is close to none.
Tailwind CSS: CSS-first configuration, design tokens in one file. Constraints in the token layer beat discipline in the markup.