OpenClaw & Hermes

Self-hosted personal agent — open alternative to Claude Desktop stack

Prev: MCP & A2A. Next: Multi-agent & KG.

Closed stack vs open stack

Hermes (Nous Hermes) — the model

Open-weight models fine-tuned for function calling and agentic chat. Solves: base Llama/Mistral often emit invalid tool JSON.

Model lineup (check Nous Research for latest)

Run locally

# Ollama
ollama pull hermes3
ollama run hermes3

# OpenAI-compatible endpoint for OpenClaw / LangChain
# OPENAI_BASE_URL=http://localhost:11434/v1
# OPENAI_API_KEY=ollama

from openai import OpenAI
client = OpenAI(base_url="http://localhost:11434/v1", api_key="ollama")
r = client.chat.completions.create(
    model="hermes3",
    messages=[{"role":"user","content":"What's 2+2?"}],
    tools=[{"type":"function","function":{"name":"calc","parameters":{"type":"object","properties":{"x":{"type":"number"}}}}}],
)
print(r.choices[0].message)

Hermes vs base model (what to demo in class)

Same 5 tool-calling prompts on Llama-3.1-base vs Hermes-3 — count valid JSON tool_calls. Hermes should win on schema adherence.

OpenClaw — the agent operating system

Always-on assistant runtime. Not a library you import — a daemon that connects channels to model + skills.

Core components

Typical config flow

1. Install OpenClaw (follow project docs for your OS).
2. Set model endpoint to Ollama Hermes or cloud fallback.
3. openclaw skills install <skill-name> from ClawHub.
4. Connect Telegram bot token in config.
5. Send message — gateway → agent loop → skill tools → reply on Telegram.

Example config (illustrative YAML)

# ~/.openclaw/config.yaml — field names vary by version; check docs
model:
  provider: openai_compatible
  base_url: http://127.0.0.1:11434/v1
  api_key: ollama
  name: hermes3

channels:
  telegram:
    enabled: true
    bot_token: "${TELEGRAM_BOT_TOKEN}"

skills:
  - name: calendar-assistant
    source: clawhub
  - name: web-search
    source: clawhub

heartbeat:
  - cron: "0 8 * * *"
    prompt: "Summarize today's calendar and unread priority emails"

Install sequence (class demo)

# Terminal 1 — model
ollama pull hermes3
ollama serve

# Terminal 2 — OpenClaw (example; use official install command)
# curl -fsSL https://openclaw.ai/install.sh | bash
openclaw init
openclaw skills search calendar
openclaw skills install calendar-assistant
openclaw channel add telegram
openclaw start

What is a Skill?

A Skill is a packaged capability — not just a prompt, not just a tool.

# Conceptual skill package structure
my-skill/
  skill.yaml          # name, description, triggers
  system_prompt.md    # behavior instructions
  tools/              # optional scripts the agent can run
    fetch_calendar.py
  config.schema.json  # user-configurable options

Same idea as Claude Skills: user installs "Calendar Assistant" — agent knows when and how to use it.

ClawHub

Community skill marketplace for OpenClaw. Browse, install, publish — like npm for agent behaviors.

How Hermes + OpenClaw work together

1. User messages on Telegram: "Summarize my calendar and email the team"
2. OpenClaw gateway receives message
3. Hermes model decides: need calendar skill + email skill
4. Skills execute tools (calendar API, SMTP or API)
5. Hermes composes summary reply → sent to Telegram
6. Heartbeat can trigger same flow at 8am daily without user prompt

OpenClaw vs LangGraph (your product decision)

Lab: OpenClaw + Hermes

1. Install Ollama + pull Hermes model.
2. Test 5 tool-calling prompts in Python (OpenAI-compatible API).
3. Install OpenClaw; point to local Hermes endpoint.
4. Install one ClawHub skill; trigger via CLI or messenger.
5. Compare: same task on base Llama vs Hermes — tool JSON validity rate.