Stack.

The tools behind my personal projects, what I use them for, and what I’m exploring next. This is a working setup, not a list of everything I’ve tried.

In use

Web applications

SvelteKit, TypeScript, and Tailwind for this site and DocLifts. PostgreSQL and Drizzle handle DocLifts’ data; Zod validates inputs at the application boundary.

Hosting and delivery

Cloudflare Pages and DNS for enoch.ai, with deployment triggered through GitHub. This site stays on Cloudflare Pages for low cost, redundancy, and less infrastructure to manage. DocLifts runs separately on an Akamai Cloud (Linode) VPS in Dallas: Ubuntu 24.04, Docker Compose, Node 24, and PostgreSQL 16. Cloudflare Tunnel serves the live Alpha at doclifts.runthe.ai; Tailscale remains available for private operations. Better Auth handles application accounts.

AI-assisted development

Hermes connects the conversation to tools, project files, and reusable skills. Claude and Claude Code are part of the planning, implementation, and review workflow. September’s DocLifts work also used Astra through OpenRouter for refactoring and features, and Codex for the historical-data import and deployment checks. Model choices can change; written requirements, human decisions, and checks against the actual result stay central.

Project context and verification

Git keeps code and design decisions together. Repository instructions give coding agents the same ground rules; automated tests and browser checks provide evidence for a handoff. For DocLifts’ history import, I restored a database locally, tested repeat imports and a scoped rollback, and compared all 14 existing application tables before and after production import. Original workout notes and labeled estimates remain attached to the imported data.

Account-scoped agent access

DocLifts’ read-only MCP server is live in Alpha. It pairs OAuth consent and revocation with account-scoped workout, program, equipment and imported-notebook reads. Bulk sets preserve historical machine context. Public protocol checks and real Muse reads passed; other clients still need individual verification.

Exploring next

Structured agent handoffs

A planned pilot with Hermes Kanban will put task state, review feedback, and verification evidence in one durable record. The goal is to resume work without reconstructing it from chat history or scattered status files. Stable project rules will remain in the repository.

Focused, access-controlled tools

I’m exploring cloud applications built around a particular person or team’s workflow, rather than a public platform. That means working through who can access a tool, what it can do, and where human approval belongs.

Have a technical constraint or a different approach worth comparing?