Browse / LLM wiki / claude-code-space
claude-code-space
efforthye
Claude Code workspace: build, deploy, and operate home-server services with self-maintained docs.
Open efforthye/claude-code-space → On the boundary of the four rules.
claude-codedeploymentdevopshome-serverhomelabllm-wikimonorepoself-hosted
Built for one agent
ships CLAUDE.md and nothing for other agents.
In its own words
A living operations wiki for building, deploying, and running real services on a home server, maintained by Claude Code. Based on Andrej Karpathy's LLM Wiki pattern: instead of re-deriving context on every question, the LLM compiles operational knowledge once and keeps it current as the system evolves. Chat history is ephemeral — this wiki is the durable memory of what exists, how it's wired, why it was built that way, how to operate it, and what has gone wrong.
You build, deploy, and decide. Claude keeps the service pages, infra map, runbooks, decisions, and incident records accurate and cross-linked.
Working with it
- New service: "create a service page for X" → Claude scaffolds the catalog entry.
- Deploy: after deploying, Claude updates the service page + deploy runbook + infra + log.
- Decision / Incident: Claude records ADRs and postmortems so context isn't lost.
- Query: ask about the system; Claude answers with citations and files good answers back.
- Lint: ask Claude to "lint the wiki" — checks stale status, missing runbooks, **leaked
secrets**, orphan pages, and open follow-ups.
Secrets never go in this repo
No passwords, API keys, tokens, private keys, or .env values in the wiki or raw/. Record pointers to where secrets live, never the values. See CLAUDE.md for the full rule.
The whole README. Open the repo ↗
Read from the repository's git tree on 2026-08-17.