Creator: Wes Roth · Published: Apr 15, 2026 · Length: approximately 43 minutes
Hermes is an open-source agent from Nous Research (the lab behind World Sim). Its core design bet is that the longer you run it, the better it gets. It keeps persistent memory across sessions, learns your projects, and auto-generates new skills which it then tries to improve through a do-learn-improve loop. Hermes is also intended to power Nous Research's agentic reinforcement learning pipeline (their Atropos project), so each user's agent doubles as a source of large-scale RL data.
Wes repeatedly references OpenClaw as the prior breakout open-source agent, built by a solo developer. He positions Hermes as the next wave: coming out of a funded, open-source lab rather than a single builder, and extending the pattern with self-improvement and messaging integrations. Hermes can also delegate tasks to Cloud Code and Codex, which matters because Anthropic's OAuth licensing lets Cloud Code be used programmatically in ways OpenClaw could not.
A distributed-training, open-source lab. Their stated position: big AI labs should not be setting the morality and ethics of models for everyone. Models should be neutral and users should decide how they behave. Hermes reflects that stance, including skills like a "god mode" skill that bundles jailbreak and prefill techniques.
Hermes ships with 74 skills on a fresh install. Categories include research (arXiv, YouTube, X/Twitter, deep research), local models (quantization, fine-tuning, Llama), media (Stable Diffusion, Segment Anything, Whisper, OpenAI text-to-speech), productivity (Notion, Obsidian, PowerPoint, Google Workspace via OAuth), developer tools (Cloud Code, Codex, GitHub review, Jupyter, PyTorch), and integrations like Polymarket and Minecraft. Skills can also be user-created or downloaded.
Wes installs Hermes on a Hostinger KVM2 VPS using their one-click Hermes template. The install is packaged as a persistent Docker container, so learned skills, memories, and configuration survive container restarts and upgrades. This matters because persistence is what makes the self-improvement loop meaningful.
You can run Hermes locally (desktop, old laptop, Mac mini, mini PC), but Nous Research itself recommends a VPS for uptime and simplicity. Wes uses Hostinger's KVM2 plan (2 vCPU cores, 8 GB RAM, 100 GB NVMe) and selects the Hermes agent one-click option at checkout.
After the VPS provisions, a configuration screen asks for:
/newbot to name your bot and generate a token.From the Hostinger browser terminal:
cd /docker/hermes-agent-XXXX (tab-complete the random suffix).docker compose exec -it hermes-agent /bin/bash to open an interactive shell inside the container.source /opt/hermes/.venv/bin/activate to activate the Python virtual environment. You will see (hermes) in the prompt once it is active.Type hermes to start the agent. Useful commands:
hermes -h shows help and flags, including --yolo to bypass dangerous-command prompts.hermes model lets you pick provider and model. Claude Opus 4.6 is the default; free options like Arc AI Trinity and NVIDIA Nemotron are also in the list.hermes setup runs the full wizard: provider, fallback credentials, text-to-speech provider (11 Labs, OpenAI, MiniMax, Mistral, Inmute, or the free edge TTS), terminal backend (local vs Docker), max tool-calling iterations (default 90, 150+ for open-ended exploration), tool progress verbosity, context compression ratio (default 0.5), and session reset policy.hermes gateway starts the messaging gateway so Telegram can reach the agent.hermes update pulls the latest version.Open Telegram, search for the bot's username (must end in bot), and hit Start (or send /start). The bot responds with a pairing code. Back in the CLI, run hermes pairing approve telegram <code> to authorize your Telegram account.
Credentials live in /opt/data/.env. List files with ls -a to see the hidden .env, then view it with cat /opt/data/.env. To edit, install nano (apt install nano if missing) and run nano /opt/data/.env. Lines starting with # are commented out; remove the # and paste the key. Save with Ctrl+X, Y, Enter. After editing the .env, restart the container with docker compose restart, then re-enter the container and virtual environment before running hermes again.
When Hermes tries a sensitive action, it prompts to Allow Once, Allow For Session, Allow Always, or Deny. The same prompt appears in Telegram with inline buttons. The security layer is named Tirith (Wes notes this is Elvish for "guard", as in Minas Tirith).
Wes argues that open-source agents are where the action is right now, and Hermes extends what OpenClaw started in two important ways. First, it comes from a lab with infrastructure, not a solo developer, so the roadmap is more ambitious (asynchronous large-scale RL, mass data generation, integration with Atropos). Second, self-improvement is built in from day one: the agent learns your projects, builds skills for recurring work, and iterates on those skills instead of starting fresh each session.
The VPS-plus-Docker packaging also lowers the bar. Wes explicitly tells viewers they do not need to be developers. Between one-click installs, AI chatbots that can explain any command or error, and Hermes' own ability to delegate to Cloud Code and Codex, the friction to running a persistent personal agent is collapsing.
cd, ls, cat, nano) and leaning on a chatbot for the rest is enough to run this.