The Claude Agent Workspace Model

25-03-2026

This is probably the most important pattern I've developed for Claude Code. The reference specification lives at Claude-Agent-Workspace-Model.

The Big Idea

A Claude Code repo doesn't have to be a software project. It can be a workspace — a structured directory that gives Claude persistent context and structured operations for any domain of work: home management, research, business ops, system administration, creative projects.

The pattern: create a version-controlled repo with a CLAUDE.md (environment briefing), .claude/commands/ (slash commands as task definitions), optionally .claude/agents/ (subagents for complex workflows), and context/ (domain knowledge). Open Claude Code in that directory and you have a purpose-built AI assistant with full context about the domain.

The Canonical Structure

workspace-name/
├── CLAUDE.md                    ← Lightweight entrypoint
├── context/
│   ├── for-agent/               ← Detailed instructions
│   └── reference/               ← Domain knowledge
├── .claude/
│   ├── commands/                ← Slash commands
│   └── agents/                  ← Subagents
├── work-log/                    ← Session outputs
├── planning/                    ← Plans and strategies
├── user-docs/                   ← Human-facing documentation
└── manifest.json                ← Agentic entry point

Three lifecycle stages: Scaffold (create the structure) → Personalise (via /onboard) → Operate (use it).

Real-World Implementations

This isn't just theory. Here are working examples across different domains:

System Administration

• Claude-LAN-Manager-0126 — Python GUI for managing LAN devices through isolated Claude spaces. Each device gets its own MCP server; Claude launched in a space sees only that device's MCPs. Effectively an AI-mediated SSH replacement.
• Claude-Proxmox-Manager-Template — 38 slash commands and 10 subagents for Proxmox server management. Fork once per machine, customise with /claude-setup.
• Claude-Home-Assistant-Manager-Template — Same pattern for Home Assistant: 20+ commands covering health monitoring, Zigbee/Z-Wave/MQTT, backup, config validation.
• Claude-Linux-Server-Manager — Generic Linux server template: Docker, XFS/RAID storage, Cloudflare Tunnels, security auditing, backups.
• Claude-Code-LAN-Manager — Lighter-weight LAN admin launchpad with ARP scanning, network topology context, and ad-hoc device management.

Desktop & Environment Management

• Claude-Code-Linux-Desktop-Slash-Commands — Library of 93+ slash commands for Linux desktop sysadmin, across 20+ categories. The source library that desktop workspaces draw from.
• Claude-Linux-Desktop-Manager — WIP: GUI wrapper exposing the slash command library as clickable tasks with visual feedback and log management.
• Claude-Conda-Manager — Workspace for managing Conda environments on an AMD ROCm workstation. Handles env audits, ROCm dependency resolution, and documentation of working package combos.

Why This Matters

The insight is that folder structure + Markdown files provide lightweight RAG without any database or vector infrastructure. Git gives you versioning. Slash commands give you structured operations. CLAUDE.md gives you persistent context. You don't need a framework or platform — just a well-organised directory.

The repo catalogues 70 workspace ideas (7 implemented, 63 conceptual). The pattern is also framework-agnostic — it can be adapted for Aider, Goose, Codex, or any agent that reads from the filesystem.

Disclaimer: This is a very opinionated approach to using Claude Code. It works brilliantly for my use cases (sysadmin, home management, research), but it's a significant departure from using Claude Code purely as a coding assistant.

Source

• Claude-Agent-Workspace-Model (reference spec)