Hermes Overview

Hermes is an alternate agent runtime for Intern. By default an Intern runs OpenClaw (Node.js gateway, plugin-based skills); switching to Hermes replaces that runtime with Hermes Agent (Python-based, persona-driven, MCP-style tooling). One runtime is active at a time — they don't run side by side.

The choice is persisted as active_agent in /root/config/config.json ("openclaw" or "hermes"). You can swap back and forth without re-pairing the device.

Where it runs

┌─ Intern device ──────────────────────────────────┐
│                                                   │
│  intern-server                                    │
│   └─ persists active_agent in config.json         │
│   └─ syncs model list every 30 min                │
│                                                   │
│  One of (never both):                             │
│   ┌───────────────────────┐   ┌────────────────┐  │
│   │  openclaw.service     │   │ hermes-gateway │  │
│   │  Node + plugins       │ ↔ │ Python + uv    │  │
│   │  ~/.openclaw/         │   │ ~/.hermes/     │  │
│   └───────────────────────┘   └────────────────┘  │
└───────────────────────────────────────────────────┘

What stays the same

The setup payload, channels, and LLM provider all come from config.json. Pairing, Wi-Fi, and the captive-portal flow don't change.

Concern	Same as OpenClaw?
Pairing (AP mode, captive portal)	yes
config.json schema (LLM key, channel tokens)	yes
Telegram / Slack / Discord channel adapters	yes
Supermemory allocation	no — Supermemory is an OpenClaw plugin and is not wired into Hermes
WhatsApp adapter	no — not supported on Hermes; setup endpoint rejects WhatsApp creds when active_agent=hermes
OTA + bootstrap	yes — both managed by bootstrap-server

What's different

Aspect	OpenClaw	Hermes
Language	Node.js	Python (uv-managed)
Active systemd unit	openclaw.service	hermes-gateway.service
Config file	/root/.openclaw/openclaw.json (JSON)	/root/.hermes/config.yaml (YAML)
Channel creds	embedded in openclaw.json plugin entries	/root/.hermes/.env (env vars)
Persona / system prompt	AGENTS.md	SOUL.md
Skill loading	~/.openclaw/skills/ registry	~/.hermes/skills/ directory
WhatsApp	supported	not supported
Install size	already on the image	~5–10 min one-time install (or pre-installed in golden image)

When you should switch

You probably want Hermes if any of:

• You want the MCP-native tooling Hermes ships with (browser automation, ffmpeg, headless chromium).
• You're testing a persona-driven (SOUL.md) agent style instead of OpenClaw's plugin-based skills.
• The Autonomous web console explicitly asks you to enroll in the Hermes program.

You probably want to stay on OpenClaw if any of:

• You rely on WhatsApp as a channel.
• You depend on Supermemory or other OpenClaw plugins.
• You want the lowest-latency cold start — OpenClaw stays in memory; Hermes has Python + node bootstrap on first run.

Where to go next

1. Manual Install over SSH — SSH into a Dev-Edition Intern and run the install steps intern-server runs internally: install the binary, register the gateway unit, write config.yaml + .env, start the gateway, persist active_agent.
2. Configuration — files on disk (config.yaml, .env, SOUL.md, skills/), the live-model sync, how to edit them safely, and how to switch back to OpenClaw.

Prerequisite: the SSH walkthrough assumes you already have shell access to the device. If you don't, see Developer Edition → SSH Access first.