Replicable Beginner Setup Guide
This is a self-hosted, beginner-friendly tutorial for setting up a governed Hermes Agent (thin orchestrator) + Grok Build CLI (sovereign brain) system on your own machine.
The steps are written generically so any reader can follow them. Replace example paths with your own (e.g. your home directory, your chosen workspace folder). The live version of this page is served from an example production deployment of this exact pattern.
Hermes is deliberately thin. It decides what and when. All real intelligence, personas, subagents, and deep work stay inside the sovereign Grok installation. "Native Grok Build TUI features first" is a core principle.
The Grok Build CLI is your permanent sovereign brain. It is installed separately from Hermes and is never replaced by it.
grok --version
Hermes acts as the persistent outer conductor (gateway, Kanban, cron, skills). Run it as a named profile (the "orchestrator" profile is a common pattern for a thin meta-layer).
curl -fsSL https://raw.githubusercontent.com/NousResearch/hermes-agent/main/scripts/install.sh | bash
hermes setup
hermes profile create orchestrator
hermes profile use orchestrator
Create a dedicated folder for governed work. This is where you will run Grok sessions with constitutional rules.
Example commands:
# Linux / macOS mkdir -p ~/grok-swarm-workspace cd ~/grok-swarm-workspace # Windows (Git Bash or PowerShell) mkdir -p C:/Projects/MyGrokSwarm cd C:/Projects/MyGrokSwarm
Inside this folder you should place (or generate) AGENTS.md and CONSTITUTION.md files that define Rule Zero and the 6 personas (see the sections below for the exact content used in production setups).
This text (or a close equivalent) belongs in your workspace CONSTITUTION.md and must be respected on every significant task:
Hermes launch helpers always inject this reminder when starting a session inside your workspace. The orchestrator itself can emit a compliant SWARM PLAN as a safe fallback.
| Persona | Primary Focus | Typical Use |
|---|---|---|
| swarm-orchestrator | Coordination, decomposition, synthesis, workflow | End-to-end ownership & delegation |
| architect | Strategy, risk, compliance, long-term vision, ADRs | Trade-off analysis & risk mgmt (often with plan mode) |
| technologist | Implementation, systems, DevOps, security, scale | Production-grade code & infra |
| creative-director | Content, design, branding, visuals, storytelling | High-impact narrative & assets |
| growth-operator | Marketing, psychology, virality, GTM, metrics | Audience & business acceleration |
| research-sage | Deep research, synthesis, fact-checking, intel | Evidence chains & truth-seeking (prefer explore type) |
In native Grok sessions these are invoked via the task tool (with persona="...") or as dedicated agent types. Always start complex work by ensuring the full set of personas is represented in the SWARM PLAN.
Use Hermes Kanban (via the orchestrator profile or dedicated workers) for decomposition, assignment, visibility, and human review gates. On platforms where the Hermes CLI is slow, direct database access to state.db is a reliable pattern.
grok -p "Your goal here. Remember Rule Zero from CONSTITUTION.md: produce a clear structured SWARM PLAN first naming the task and assigning the 6 personas. Wait for explicit approval before any tool use. Operate inside /path/to/your-grokswarm-workspace under full constitutional governance." \ --cwd "/path/to/your-grokswarm-workspace" \ --max-turns 25 \ --no-memory \ --output-format plain \ --no-alt-screen
Always set the correct --cwd to your workspace. Prefer --prompt-file for long constitutional reminders. Use --permission-mode plan for high-stakes architectural work.
grok -p "Output the exact text of Rule Zero from CONSTITUTION.md and the current SWARM PLAN template from AGENTS.md. Do not create any plan or do any work." --cwd "/path/to/your-grokswarm-workspace" --max-turns 5 --no-memory --output-format plain