A year ago my “AI setup” was a chat window. Today it’s a team: one orchestrator, five specialists, a shared memory, and a monitoring net over five machines — all running on hardware I already own, for effectively zero dollars a month.

This is the architecture, and why it’s worth copying.

The team

The core is an OpenClaw gateway on my Razer Blade. Aura is the orchestrator — it routes work, keeps context, and owns the memory. Under it, five agents with separate models and workspaces:

• The Architect (coder) — backend, APIs, databases, services
• The Artisan (creative) — frontend, design, content
• The Vizier (researcher) — planning, market and chain analysis
• The Clerk (ops) — git, cron, deploys, config hygiene
• The Diagnostician — debugging, incidents, failures

Delegation is by protocol, not vibes: one task, one owner, verify before reporting. Only the Architect and the Diagnostician touch code. Everyone else stays in their lane.

Memory and knowledge

Everything lands in an Obsidian vault backed by git — daily notes, plans, audits, a skills registry, and a mirror of the long-term memory file. The vault is the source of truth; the chat is just the front door.

Backups are restic, nightly, to a second machine, keeping 14 days of dailies, 8 weeklies, and 6 monthlies. Losing the vault means losing the team’s memory — so it gets treated like production.

Monitoring

Five hosts are watched with read-only SSH health checks every morning: the trading box with its API and dashboards, the laptop running the rental manager, the Bitcoin Gold production VPS, an external vantage point, and the backup target. Results land in Telegram. If something dies, I know before my coffee does.

The bounty pipeline

The same team hunts bug bounties. A scanner runs scheduled passes against HackerOne targets on cloud models, the Diagnostician validates the shortlist and builds working PoCs, and the reports go out under my researcher alias. Cheap models do the volume; the strong model does the judgment.

Cost

The gateway, the agents, and the local models run on hardware I already had. Long-running work moves to free-tier cloud models. The whole stack costs effectively zero — the expensive part was the discipline: delegation protocols, a real memory, and backups that actually run.

The tools change monthly. The architecture is the point.