Complete guide to mastering Hermes Agent
Why I switched from OpenClaw to Hermes
I’ve spent hundreds of hours inside Openclaw. I love it. But I still wake up to an agent that forgot everything I told it yesterday, I went looking for something better.
I found Hermes. Then I realized the real answer wasn’t one or the other — it was both.
Here’s the full breakdown of how I installed Hermes, connected it to my Telegram, Gmail, Calendar, Apple Health, Threads, and Openclaw, and turned it into a personal LLM wiki that actually improves over time.
---
The Problem
Openclaw is great. But as a daily driver? It kept failing me in three ways.
First, memory Every morning I’d re-paste the same context. My business. My tone. My calendar. My health goals. The agent had no persistent wiki of me.
Second, continuous improvement. I want my AI agent to get smarter over time. Yet I didn’t see Openclaw do this, I needed to prompt it to improve it.
Third, instability. Part of it is my fault, I’m running on a tiny Jetson Nano. It crashed 6 times a day especially when I needed to use the browser. Openclaw is HUGE. It eats up alot of memory and space.
Andrej Karpathy recently published a piece called “LLM Wiki” — the idea that an AI agent should incrementally build and maintain a persistent wiki about you that gets better over time. That’s exactly what I wanted. Openclaw couldn’t do it. Hermes can.
---
The Solution — Install and Import
One command to install. When Hermes detected my existing Openclaw setup, it asked if I wanted to import my souls, my user profile, and my existing skills. I hit yes. Everything carried over. No re-setup.
Then it asked quick setup or full setup. I picked quick and went straight to picking my LLM.
---
The Solution — Picking Your LLM (This Is Where You Save Money)
Three options, each with real trade-offs.
Option 1: OpenAI Codex subscription.
This is what I use. $20/month, already paid for, and — critically — it’s the *only* major provider that doesn’t ban you for running an agent on top of your subscription plan. Anthropic, Gemini, and Kimi will shut you down. Codex won’t. Even better, Hermes picked up my Codex login from my machine without making me re-authenticate through a browser URL like Openclaw does.
Option 2: OpenRouter with auto-routing.
This is the power-user move. One API key, access to every model, and Hermes routes your cheap tasks to free models and your complex tasks to premium ones. As of April 13, OpenRouter has a free model called **Elephant** that performs surprisingly well. Put in $10 minimum and you unlock 1,000 free requests per day.
Option 3: Local AI via Ollama.
Free. Zero tokens. One command — `ollama launch hermes` — and it picks up your local model automatically. I tried Gemma 3 4B first (mistake — no tool calling), then switched to Qwen 3.5 9B and it ran fine on my MacBook. The catch: your local machine needs the horsepower, and less powerful models mean weaker output.
---
Telegram, Groups, and Topics
Run `hermes gateway setup`, pick Telegram, open BotFather, type `/newbot`, name it, copy the token back, then grab your Telegram ID from `@userinfobot` (watch for scam imitators — it has to be exactly `userinfobot`).
Here’s the trick nobody mentions. By default your bot won’t respond in group chats. You have to go back to BotFather → your bot → Settings → Group Privacy → **turn it off.** Then remove the bot from the group and add it back. Now it responds.
The killer feature is topics Enable topics in your Telegram group and you can create one per workflow — Social Media, App Development, Analytics, Health. Your bot runs in all of them but the conversations stay cleanly separated. No more one giant DM thread with 40 unrelated automations smashed together.
---
Dashboard and Cron Jobs
Type `hermes dashboard`, open `127.0.0.1:1919`, and you get a web UI. Recent sessions, connected platforms, token usage, error logs, cron jobs, skills toggles, config file editor.
Cron jobs are the heart of what makes Hermes different from Openclaw. I created one called “Weekly App Sales Report” — grabs my App Store data every Monday at 8:30am and delivers it to Telegram. The cron syntax is standard: minute, hour, day-of-month, month, day-of-week.
I also built a heartbeat — the feature Openclaw’s founder talked about in his TED talk. His command was simply “surprise me.” I created a cron that runs every hour and checks on what I’m working on, my health stats, or ways to improve my business. Openclaw can’t do this natively. Hermes can.
---
Souls, Agents, and User Memory
This is the part that most tutorials skip. To make Hermes actually feel like *your* agent, you edit three markdown files:
Souls.md — personality. Mine says “concise technical expert, no fluff, just tactics.” Find it with Command+Shift+Period to show hidden files, open the Hermes folder, edit the file.
Agents.md — the brain. Specific instructions. My coding style. My LinkedIn post format. What to always do, what to avoid. Rules that apply every time.
User.md — memory about me. My name, my wife, my city, my job. The stuff Openclaw forgot every morning. Hermes reads this every session.
---
Real Use Cases
I connected my own Apple Health app (Health Data AI Analyzer) via API. Hermes wrote Python code on the fly, pulled my sleep data, found I average 7.59 hours, high of 9.8, low of 5.37. Then I told it to save it as a skill and run it as a morning cron. Now it reports daily without re-writing code.
I connected Threads. Before Hermes, I needed a flaky Openclaw browser extension that scraped my screen. With Hermes, I exported my cookies via a browser extension, dropped them in, and pulled 34 posts of analytics in one command — likes, replies, reposts, and a post-by-post performance analysis. Hermes is dramatically better than Openclaw at browser automation. I open-sourced the skill.
I connected Gmail and Calendar through Google Cloud Console — create project, enable Gmail API and Calendar API, download OAuth JSON, drag it into Hermes, authorize. Now it reads my week and cross-references my sleep against my schedule.
---
Provider Routing (The Secret Money-Saver)
Through OpenRouter, Hermes routes based on price, throughput, or latency. Simple UI fix? Cheap model. Planning a complex feature? Opus 4.7. Just type `mode cheap` or `mode plan` and it switches.
Set up fallback providers too. Anthropic’s servers have gone down before. With OpenRouter, your fallback can be OpenAI, then another, then a free model. You can even route certain tasks to free Elephant and burn your daily free requests before switching to paid.
---
Why I Still Use OpenClaw
Here’s the part most Hermes videos get wrong. It’s not Hermes or Openclaw. It’s Hermes orchestrating Openclaw.
Hermes is better at: memory, persistent context, cron jobs, browser automation, quick tasks, one-shot workflows, personal daily driver use.
OpenClaw is better at: multi-agent workflows, deep research, long complex tasks, heavy execution.
The setup that actually works:
Hermes is the CEO. Openclaw is the senior engineer. Hermes handles daily life and delegates the hard stuff. I install both on the same machine, point them both at the same Obsidian vault on my UGreen NAS, and they share the exact same knowledge base.
All my souls, agents, memory, emails, calendar data, business notes — one vault, two agents, both reading and writing the same markdown files. Even if I switch machines, everything is there.
To wire this up, I give Hermes a prompt like: “You have access to Openclaw installed locally. Use Hermes for simple reasoning, quick tasks, coordination. Use Openclaw for multi-step workflows, deep research, long-running tasks. When the task is complex, delegate to Openclaw, wait for the response, and return a clean answer.”
I tested it. Hermes found Openclaw. It delegated. It returned. It works.
---
The Results
Before: $64/week on Claude API, zero persistent memory, context re-pasted every morning, no scheduled tasks, flaky browser automation.
After: $20/month OpenAI Codex subscription, persistent LLM wiki that improves daily, cron jobs running 24/7, reliable browser scraping, orchestrated two-agent system sharing one vault.
My mornings now start with a Telegram message: sleep score, calendar, flagged emails, a proactive “surprise me” insight. **I stopped using AI like a chat. I started running a system.**
---
Tools/Resources
- Hermes agent - Free, one-command install
- OpenAI Codex subscription ($20/month) - Only major provider that won’t ban you for agent use
- OpenRouter - Free tier via Elephant, $10 unlocks 1,000 requests/day
- Ollama - Free local model hosting
- DigitalOcean droplet via OpenRouter Spawn ($18/month) - For 24/7 VPS, 2 CPU + 4GB RAM minimum
- Obsidian - Free, the shared wiki layer
- UGreen NAS - Central vault storage across machines
- BotFather + @userinfobot - Free Telegram setup
- Google Cloud Console - Free OAuth for Gmail/Calendar
- Total: $20-58/month depending on setup
---
Your Next Steps
1. Install Hermes with one command and import your existing Openclaw config
2. Connect OpenAI Codex as your LLM so you stop paying per-token
3. Set up Telegram with BotFather, enable topics for clean workflow separation
4. Edit your Souls.md, Agents.md, and User.md — give your agent a personality, rules, and memory of you
5. Create one cron job — a morning heartbeat that surprises you with something useful
6. Install Openclaw alongside Hermes, point both at one Obsidian vault, and let Hermes delegate the hard work
---
Let’s Talk Personal AI Systems
Are you still using AI like a chat window? What’s the one workflow you’ve re-prompted more than 10 times that should be a cron job or a skill by now? And if you’ve tried both Hermes and Openclaw, which one ended up as your daily driver?
Drop a reply and let me know.
---
See you next week,
Keith
---
P.S. - Hermes isn’t something you install and play with. It’s the center of your personal AI system. Once it’s connected to your email, calendar, health data, social accounts, and Openclaw, it stops feeling like a chatbot and starts feeling like staff.
---
Weekly Resource Drop
If you want my exact Souls.md, Agents.md, the provider routing prompt, the Threads analytics skill I open-sourced, and the Openclaw delegation prompt — join my substack email. This week I’m dropping the full config I use to run both agents against one shared Obsidian vault.
---
*Good? Ok? Bad? Hit reply and let me know...*
oh sorry i have a question, i host on VPS, i saw the webUI port from some tutorials is 9119, what should be the actual port?