hermes-setup

My Mac mini deployment of NousResearch/hermes-agent — "the agent that grows with you."

This is the sanitized public scaffolding for how I run hermes-agent unattended on a Mac mini at home: launchd timers for deterministic shell, OpenClaw for scheduled Claude agent invocations, and the glue between them.

It's the unattended companion to my interactive Claude Code setup in claude-code-starter.

What this is (and isn't)

Is: patterns, scaffolding, install scripts, sanitized config templates, plist examples, the operating model.

Isn't: my actual cron payloads, real hostnames, real Telegram chat IDs, case-specific scripts, or anything that names real people. That stays in a private hermes-backup repo behind git-crypt. See claude-code-starter's Privacy section for the same firewall pattern.

Architecture

┌─────────────────────────────────────────────────────────┐
│  Mac mini (always on)                                    │
│                                                          │
│  ┌──────────────────────────┐                           │
│  │ hermes-agent             │  ← grows with use         │
│  │ (NousResearch)           │     Anthropic API direct  │
│  └──────────────────────────┘                           │
│              ▲                                           │
│              │ invoked by                                │
│              │                                           │
│  ┌─────────────────────┐   ┌─────────────────────────┐  │
│  │ OpenClaw            │   │ launchd                 │  │
│  │ (LLM scheduling)    │   │ (shell scheduling)      │  │
│  │ "agent cron"        │   │ "deterministic cron"    │  │
│  └─────────────────────┘   └─────────────────────────┘  │
│              │                       │                   │
│              └───────────┬───────────┘                   │
│                          │                               │
│                  ~/.claude/                              │
│              (skills, agents, hooks,                     │
│               shared with the laptop)                    │
└─────────────────────────────────────────────────────────┘
                          │
                          │ Tailscale
                          ▼
                  Laptop (interactive)
                  claude-code-starter

Operating model (two rules)

1. openclaw cron is ONLY for LLM work. Agent turns, browsing, summarization, code review on a PR queue. Anything that needs the model.
2. Pure shell work goes in launchd plists. git pull, rsync, curl, systemctl restart. Burning agent tokens on deterministic shell is pure waste.

This split is what makes the combined cost sustainable.

Layout

hermes-setup/
├── README.md                       # you are here
├── LICENSE                          # MIT
├── install.sh                       # idempotent installer; refuses to run with leftover <PLACEHOLDER>s
├── .gitignore                       # blocks .env, secrets, runtime state
│
├── hermes/                          # sanitized ~/.hermes/ templates
│   ├── README.md
│   ├── config.example.yaml          # ~/.hermes/config.yaml template
│   └── SOUL.example.md              # ~/.hermes/SOUL.md template (operator personality)
│
├── launchd/                         # macOS LaunchAgent plists
│   ├── README.md                    # launchd vs openclaw cron, gotchas, useful commands
│   ├── ai.hermes.gateway.plist      # hermes-cli gateway daemon (real, sanitized)
│   └── com.example.daily-backup.plist  # sample deterministic shell job
│
├── openclaw/                        # scheduling patterns
│   ├── README.md
│   ├── openclaw.example.json        # minimal ~/.openclaw/openclaw.json template
│   └── cron-examples.md             # 3 sanitized scheduled-prompt patterns
│
└── tools/                           # generic glue
    ├── README.md
    └── hermes-doctor.sh             # health check; supports --remote <ssh-host>

Install

git clone https://github.com/barkleesanders/hermes-setup.git
cd hermes-setup

# Preview what install.sh would do without writing anything:
./install.sh --dry-run

# Run for real:
./install.sh

The installer refuses to overwrite existing ~/.hermes/config.yaml or ~/Library/LaunchAgents/ai.hermes.gateway.plist, so re-running it is safe.

After install:

hermes-cli auth login                          # if you haven't already
$EDITOR ~/.hermes/config.yaml                  # tune the kawaii/cost knobs
$EDITOR ~/.hermes/SOUL.md                      # set operator personality
./tools/hermes-doctor.sh                       # should be all green

What it actually runs

Want to see what kinds of crons, tools, and skills this setup exercises in practice? See WHAT-IT-RUNS.md for a sanitized bucket-level inventory: cron categories, tool families, skill usage, and the deliberate boundaries (no money-moving, no auto-replies to outside parties, etc.).

Related

• NousResearch/hermes-agent — the upstream agent runtime this repo deploys.
• claude-code-starter — the interactive ~/.claude/ config that hermes-agent shares. Same skills, agents, hooks; different invocation context.
• OpenClawBS — the generic Mac mini scheduling reference architecture. hermes-setup is one concrete instance of that pattern.
• OpenClaw — the scheduler used by both.

License

MIT. See LICENSE.