Your AI coding agent from the earlier lesson answers only when you sit down and ask it something. This lesson covers a different kind of assistant: one that stays on all the time, remembers what you told it last week, and answers you on Telegram or WhatsApp while you are at lunch. That is a real unlock for a small business or nonprofit: a drafted reply waiting in your inbox, a reminder that fires itself, research pulled together overnight, a file summarized while you sleep. It is also a bigger risk than anything you have set up so far, and this session names that risk before you install anything.
The project is Hermes Agent, built by Nous Research (the lab behind the open Hermes model family), open source under the MIT license. Repo: github.com/NousResearch/hermes-agent. It installs on a computer you own and runs as a background process, "the agent that grows with you," in its own words. Unlike a chatbot in a browser tab, it keeps memory across every conversation: notes about itself, a profile about you, and a searchable log of what you have told it. You reach it from a terminal or a chat app, Telegram and WhatsApp both fully supported, alongside many others (Discord, Slack, Signal, email, SMS), through one background "gateway" process.
For a business or nonprofit, that unlocks the work that happens between meetings: a draft reply waiting in your inbox, a reminder that fires, research assembled overnight, a file worked on while you are out. It does not replace you. It works while you are somewhere else.
Your coding agent is a tool you drive: you open a terminal, type, it acts, you close the window. Hermes Agent is closer to an employee with a phone: it runs continuously whether or not you are looking, and it can act on messages that arrive from the outside world, not just instructions you typed yourself. That is the whole appeal, and exactly where the risk starts.
Here is the honest version, close to the project's own security documentation. You are handing this agent two things at once: your entire computer (its default setup gives it "the same filesystem access as your user account," nothing sandboxed unless you configure it) and a bot that reads whatever strangers send it. An independent researcher who tested Hermes as a Telegram assistant found the default setup ships a terminal, web browsing, and file tools right alongside the channel that receives untrusted messages, and concluded this turns an ordinary prompt-injection trick into a path to real command execution or stolen data. Approval prompts help, but they fire at the wrong moment: a hidden instruction can act before you ever see a prompt to approve.
The project's own guidance agrees, and boils down to three rules worth writing down before you touch anything:
Next session, only ever run the installer from hermes-agent.nousresearch.com or github.com/NousResearch/hermes-agent. Search results turn up several look-alike domains claiming to be Hermes Agent resources; none belong to Nous Research. Only hermes-agent.nousresearch.com is linked from the real Nous Research site and listed as the project's own homepage. One look-alike checked out fine when tested, but that proves nothing about the others, or tomorrow. Installers run with your full permissions the moment you paste them in; the "verify before you run" habit applies doubly here.
Hermes Agent runs on Linux, macOS, WSL2, Android (Termux), or Windows; Git is required, and the installer pulls in the rest itself. You will pick a model provider at setup: OpenAI, Anthropic, Google, dozens of others, or a fully local model. The project publishes no official minimum spec or subscription price; its docs describe it running anywhere from a small VPS to a full GPU cluster, so treat any dollar figure you see elsewhere as an estimate, not a promise from Nous Research.
Session 2 sets it up on your spare machine and connects a model. Session 3 covers memory and gets Telegram working end to end. Session 4 locks it down for real: manual approval on risky actions, an allow-list of who can message it, checkpoints you can roll back, and the update habit.
You have decided which spare machine you will use, or that this is not for you yet. Both are a fine outcome for today. And you have written the three rules on a sticky note: spare machine, your own IDs only, updates on.
Template included: Members get the readiness checklist: which spare machine you will use, the accounts to create, the keys to keep separate from your main accounts, the install-domain rule, and the three rules as a printable card. Find it in your member dashboard under Templates.
Short on time? Alliance Optimal will implement this end to end, wired into your stack, and hand you the keys.