The Hermes Agent Field Guide: The Open-Source AI That Writes Its Own Skills

Hermes Agent is Nous Research's open-source AI agent — MIT-licensed, and one of the faster-growing agent projects on GitHub. From their own docs: it's "the only agent with a built-in learning loop." Most agents forget everything between sessions and run on someone else's servers. Hermes does neither. It runs on hardware you control — your laptop, a cheap VPS, or serverless like Daytona/Modal that hibernates when idle. It remembers across sessions. And it writes its own skills from the work you actually do. You don't rent this assistant. You grow it.

• Open-source, MIT-licensed, from Nous Research (github.com/NousResearch/hermes-agent)
• Runs on YOUR infra — laptop, VPS, or serverless; a dedicated server is optional
• Persistent memory across sessions (not a fresh, forgetful chat each time)
• It writes and updates its own skills — the self-improving loop is the whole point

This is the part people overcomplicate. On macOS, Linux, or WSL2 it's one line, and the installer pulls in everything it needs — uv, Python 3.11, Node 22, ripgrep, ffmpeg. The only thing you install first is Git. Then one setup command wires up a model and the tools. Want a desktop app instead? There's an installer at hermes-agent.nousresearch.com/desktop.

1. Install (macOS/Linux/WSL2): curl -fsSL https://hermes-agent.nousresearch.com/install.sh | bash
2. Install (Windows PowerShell): iex (irm https://hermes-agent.nousresearch.com/install.ps1)
3. Reload your shell: source ~/.zshrc (or source ~/.bashrc)
4. One-command setup: hermes setup --portal — OAuth login, picks a model, turns on the tool gateway
5. Start it: hermes --tui (the modern interface; plain 'hermes' also works)

Read these as the mental model behind the agent, not a to-do list. A stateless agent is only as good as what you re-explain to it every time. Hermes is built to skip that — it keeps the context, and it keeps getting sharper on its own.

• Memory — durable facts and preferences carried across every session, with full-text recall of past conversations
• Skills — reusable 'recipes' (a markdown file + YAML front-matter); Hermes writes new ones from your real work and patches them when you give feedback
• Tools — 60+ built in (web, files, terminal, code execution, image/voice/vision, browser control), plus anything you add via MCP servers
• Automations — cron schedules that wake a fresh agent session at the right time and report back
• Channels — it lives where you already are: Telegram, Discord, Slack, WhatsApp, Signal, email and more

This is the part that pulls people in. You don't sit at a terminal. Connect a chat channel and talk to it from your phone. It wakes instantly, runs the task, and messages you back — voice or text. One command sets it up.

1. Connect a channel: hermes gateway setup — then pick Telegram (or Discord/Slack/WhatsApp/Signal/email)
2. Pick your model brain: hermes model — Nous Portal (zero-config, 300+ models), or OpenAI / Anthropic / OpenRouter with your own key
3. Note the one real requirement: the model must support at least 64,000 tokens of context
4. Stuck? run hermes doctor to check your setup; hermes -c resumes your last session

Skip the big ambitious build on day one. Get it talking first. Then hand it one thing you actually repeat and let it save a skill for it. The next day, ask it to run that skill — it'll have remembered, and you can tell it to do the job better.

1. Install + hermes setup --portal, then hermes --tui
2. Connect Telegram with hermes gateway setup so you can use it from your phone
3. Give it one real task you repeat (a daily summary, a research routine), and tell it to save a skill for it
4. Come back tomorrow and ask it to run that skill — notice it remembered, and ask it to improve it