The author built a sandboxed LLM development environment on a dedicated eBay server where Claude can autonomously write, test, deploy, and iterate on web applications without touching the author's main systems. The setup uses Coolify (self-hosted PaaS), Forgejo (Git + CI), Hermes (LLM agent), and Tailscale networking to create an isolated "software factory" accessible only via home network tunnel. The isolation strategy is practical but honest: rather than full sandboxing, the author accepts that the LLM can still destroy this server and its databases—the sacrifice is acceptable because it can't reach systems that matter. The demo shows Claude building a calorie-tracking SvelteKit app (Postgres, Tailwind, Docker) from a single natural-language prompt, committing code, fixing test failures, and deploying it to a vanity HTTPS URL, all autonomously. The source document is well-written and grounded in actual implementation choices (DNS-01 cert generation, Porkbun API integration, Coolify Docker Compose setup) rather than vaporware. One sharp question: what happens in practice when the agent needs to make outbound API calls (email, SMS, payment providers)—does each credential get its own machine, or does the author accept broad token blast-radius as a solved problem?