Hermes Agent

What it is

Hermes Agent is an AI agent from Nous Research that you run yourself. Its README calls it self-improving. It creates skills from experience, improves them as it uses them, keeps memory across sessions and searches its past conversations. You talk to it in a terminal, in the Hermes Desktop app, or from Telegram, Discord, Slack, WhatsApp or Signal through one gateway process. It works with the model provider you choose, and one command, hermes model, switches it.

1. You talk to it in a terminal, the desktop app or a chat app
2. It runs tools in a backend you pick, from local to Docker
3. It keeps memory and writes a skill after a complex task
4. Later sessions reuse the skill and improve it

Who it suits

People who want one assistant that keeps running, remembers them and can be reached from a phone while it works on a server. Developers get a terminal agent that can work in isolated git worktrees. People leaving OpenClaw, Claude Code, Codex CLI or Cursor can import their settings, memories and skills. Skip it if you only need a single chat box, or if your model has less than 64,000 tokens of context, which Hermes rejects at startup.

Where to use it

The docs list these places.

• Hermes Desktop on macOS (Apple silicon only) and Windows
• The command line on Linux, macOS on Apple silicon and WSL2, and natively in Windows PowerShell
• Telegram, Discord, Slack, WhatsApp, Signal and email through the gateway, plus more platforms its docs list
• Terminal backends: local, Docker, SSH, Modal, Daytona and Vercel Sandbox
• Android through a Termux package, which the docs mark as broken right now

How to set it up

Get it saves a zip of Mindket’s setup notes in your downloads folder, with the maker’s steps for each platform. The Quickstart recommends the Hermes Desktop installer on macOS and Windows. On Linux, macOS and WSL2 there is a one-line installer.

1. Run the installer for your system.
2. Reload your shell.
3. Run hermes setup, or hermes model to choose a provider.
4. Run hermes and send a first message.

curl -fsSL https://hermes-agent.nousresearch.com/install.sh | bash

On Windows, run this in PowerShell instead:

iex (irm https://hermes-agent.nousresearch.com/install.ps1)

On Linux, macOS and WSL2, reload your shell and start:

source ~/.bashrc
hermes

On Windows, run hermes.

How to use it

Type hermes to start a chat, or hermes --tui for the newer terminal interface. Slash commands such as /model, /new and /retry work in the TUI (hermes --tui) and in chat apps. The classic CLI takes a shorter list that includes /model, /compress and /stop. Type a new message while it works to interrupt and redirect it.

Example

You run Hermes on a small server and want a news summary on your phone. You run hermes gateway setup, pick Telegram and give it the bot token from BotFather and your Telegram user ID. Then you start the gateway and send the request to your bot in Telegram, in plain words. The Cron page uses this exact request and says Hermes creates the job with its cronjob_manage tool.

Every morning at 9am, check Hacker News for AI news and send me a summary on Telegram.

The job then runs on its schedule, and the Cron page says results can be delivered back to the chat where you asked.

For newcomers and experts

If you are new. A skill is a Markdown file that teaches the agent one task. Hermes reads each skill’s short description and loads the full text only when a task needs it. Start with hermes setup and one plain chat. The Quickstart says to get that working before you add the gateway, scheduled jobs or skills. If something feels off, run hermes doctor first.

For experts. Secrets go to ~/.hermes/.env and other settings to ~/.hermes/config.yaml, and hermes config set writes each value to the right file. Blank Slate setup starts with only the provider, file operations and the terminal on. Terminal backends switch with hermes config set terminal.backend. A cron job runs on its own pinned model, else on cron.model, else on whatever hermes model is set to at the moment it fires.

An MCP server goes under mcp_servers in config.yaml, and each one can be filtered to the tools you allow. The MCP page gives this entry:

mcp_servers:
  filesystem:
    command: "npx"
    args: ["-y", "@modelcontextprotocol/server-filesystem", "/home/user/projects"]

Skills install from the Skills Hub after a security scan, and an install names its source, as in hermes skills install openai/skills/k8s. Imports from Claude Code, Codex CLI and Cursor can be re-run later with hermes import-agent --sync.

Good to know

Hermes runs commands for you and keeps keys on your machine, so a few points matter before you start.

• Installers: the one-line installers download and run a script from the maker’s site
• Providers: Hermes needs the model provider you pick, which can be a hosted API or your own endpoint such as LM Studio or Ollama. Secrets and tokens sit in ~/.hermes/.env
• Approvals: Hermes checks a command against a list of dangerous patterns before it runs. In the default smart mode an auxiliary model judges the risk. Turning approvals off, or typing /yolo, skips the prompts
• Gateway: a Telegram bot takes an allowlist of user IDs, and its token must stay secret because anyone holding it can control the bot
• Imports: API keys and credentials are never imported, except OpenClaw’s allowlisted API keys and bot tokens. hermes setup passes --migrate-secrets for OpenClaw, and by hand they come over only with --migrate-secrets
• Windows: an antivirus may flag uv.exe, which the README calls a false positive and tells you to verify
• Android: the Termux package is marked broken in the docs
• Removal: a source install is removed with hermes uninstall. Review hermes uninstall --dry-run first, because --full and --data delete data. The Desktop app is removed in Windows Settings → Apps → Installed apps, or on macOS by quitting Hermes and moving the app to the Trash
• License: MIT. The LICENSE names Nous Research