Published: 2026-04-13

Hermes as Brain, OpenClaw as Arms: The Chief-of-Staff Multi-Agent Setup

Chapters / key moments (click to jump — plays here on the page)

Craig Hewitt demonstrates a practical Hermes + OpenClaw multi-agent architecture: Hermes acts as the always-on chief of staff (memory, orchestration, long-term context) while named OpenClaw sub-agents — Gary for email copywriting, Patty for YouTube strategy — handle specialized execution. The entire system runs 24/7 on a $9/month Hostinger VPS, deployed in minutes via Docker. Craig adapted this from Ryan Carson's open-source implementation.

Source video

"Hermes Just Solved the Biggest Problem With OpenClaw" by Craig Hewitt — Watch on YouTube →

Key Takeaways

Hermes solves OpenClaw's memory problem. OpenClaw's context resets per session and can't compound knowledge over time. Hermes was designed for long-running persistent memory — routing all AI work through it means your agent gets smarter with every interaction rather than starting fresh each time.
OpenClaw sub-agents get specialized skills from GitHub repos. Point a new OpenClaw instance to a skill repo (e.g., Corey Haynes's marketing-skills repo) and it becomes a domain expert in seconds — no manual configuration, just one instruction to "get this repo and implement these skills."
Deploy both agents on a VPS for 24/7 availability. A Hostinger KVM 2 plan ($9/month) handles both Hermes and OpenClaw in Docker containers. Power outages or travel no longer kill your agent setup. One-click deploys through the Hostinger dashboard — no terminal required.
The "Chief of Staff" system prompt is the key architectural piece. Craig uses a free GitHub repo (linked in description) that gives Hermes a structured role: personal priorities, delegation patterns, skill awareness, and communication preferences. Without this, Hermes is a blank slate.
OpenClaw is never user-facing in this setup — it's a sub-agent only. Users communicate with Hermes via Telegram. Hermes decides when to invoke an OpenClaw sub-agent and packages the brief it sends. OpenClaw focuses entirely on execution, not conversation.

How the Live Demo Works

Craig shows a real end-to-end run: he asks Hermes (via Telegram) to call Gary, his email/copywriting OpenClaw sub-agent, and create an email newsletter about smart model switching for AI agents. Here's the sequence:

Hermes pulls its chief-of-staff skill and identifies that this request requires the Gary sub-agent.
Hermes researches model-switching best practices using its web search tools (Minimax M2.7 via OpenRouter).
Hermes prepares a structured brief with audience, objective, key points, and tone instructions.
Hermes calls the Gary OpenClaw sub-agent with the brief via an "agent call."
Gary returns the completed newsletter copy to Hermes.
Hermes presents the output to Craig in Telegram with full context about what was done.

Total time from "call Gary" to newsletter draft: a few minutes. The compounding value: Hermes's knowledge of Craig's audience, preferences, and previous campaigns informs the brief — knowledge that Gary doesn't have and doesn't need.

Why This Architecture Works

The video articulates a key principle: route as much work as possible through one place (Hermes) because compounding context is where the real leverage lives. Every piece of work Hermes sees — your emails, your creative decisions, your feedback on Gary's copy — becomes context for the next task. OpenClaw sub-agents are powerful but stateless between calls. Hermes provides the continuity layer.

Think of it this way: Hermes is your COO. Gary is a specialist contractor. The contractor is excellent at their job but has no institutional knowledge. Your COO briefs them with everything they need, evaluates their output, and files away what was learned. That's the architecture Craig is running.

Related on OpenClawDatabase

Hermes Agent Guide — full guide to Hermes setup, memory, tasks, and MCP tools
OpenClaw Skills Guide — how to give OpenClaw sub-agents specialized capabilities
OpenClaw + Hermes Multi-Agent Workflow — Alex Finn's complementary take on the same architecture

What you can actually set up from this

Extracted from the video's own transcript — the specifics the original summary left out.

Reproducible steps

The topology: one brain, several hands
Hermes runs as the persistent brain and chief of staff; OpenClaw instances run as dedicated single-purpose subagents that do the work. Everything sits on one VPS so it is always on. The argument for routing as much work as possible through one place is that the context compounds there — the brain gets smarter over time, which a scattering of separate sessions never does.
Give each subagent a name and a single job
The two shown are an email/copywriting agent and a YouTube strategy agent, each named after a practitioner in that field. The delegation chain in the demo is explicit: Hermes recognises the request needs the email agent, does the research itself, writes a brief, sends the brief to the OpenClaw agent, and the result comes back into Hermes — so the context stays in one place even though the work did not.
Skills are SOPs, and that is the right mental model
A skill is written the way you would onboard a new employee: here is the job, here are the steps, here is the definition of done. The YouTube agent carries more than a dozen — audience research, channel audit, channel strategy, descriptions, hook writing. One skill quoted in full covers writing the first 30 seconds of a video and creating retention hooks.
Bootstrap from published skill repositories
The setup path is to point a blank agent at a public repo URL and tell it to implement that set of skills. Two are named: a chief-of-staff repo for the Hermes brain, and a general marketing-skills repo for building a marketing agent from nothing.
Run it on a VPS with Docker, for a specific reason
The reason given is concrete rather than theoretical: the presenter was travelling when the power went out at home, his local machine shut down, and he lost access to every agent and all their context for three days. Docker is used so an agent cannot take down the whole server.

Gotchas

The video is sponsored by the VPS host and includes a discount code. Nothing in the architecture depends on that host — any VPS with SSH and Docker works. The plan tier used is a ~$9/month 2-core instance, which is a useful floor to know.
The model split is deliberate: a low-cost model on the Hermes brain (MiniMax M2.7 via OpenRouter in the demo) with the expensive work delegated. Cheap orchestration plus targeted spend is the pattern, not cheap everywhere.
Hostinger's one-click deploy pre-bundles some API keys with the plan; the setup wizard asks for OpenRouter, Anthropic and OpenAI keys and can be left blank at install time and filled in later.
The setup is adapted from a published implementation by another practitioner and is free to download — the video is not selling the architecture, only the hosting.

← Back to News digest · See also: Hermes guide · OpenClaw guide