Zero to Agent: A Complete Setup Guide for Hermes Agent

Nous Research's open-source AI agent Hermes Agent requires deployment on a virtual private server (VPS) to run continuously, including while users sleep, as the desktop version stops when the laptop closes. The agent features a self-improving learning loop, procedural memory via skills, and durable user memory, but users must monitor model provider balances to avoid silent failures. The setup guide explains key concepts such as VPS, containers, API keys, tools, skills, and memory.

27 min read A step-by-step walkthrough for setting up Hermes Agent, the open-source AI agent from Nous Research. Table of Contents What Hermes Agent Actually Is Start with what it isn’t. It isn’t a chat window you visit when you have a question. A chatbot is reactive: it waits for you, answers, and forgets the moment you close the tab. Hermes Agent is closer to hiring someone. It stays switched on, checks the tools it has access to, keeps notes on what you’ve told it, and can act on a schedule without you sitting there prompting it turn by turn. The project’s own description calls it a self-improving agent with a built-in learning loop: it creates its own procedures from experience, nudges itself to save what it learns, searches back through its own past conversations when it needs context you didn’t repeat, and builds up a working model of who you are the longer you use it. It’s maintained by Nous Research, and because the source is public under an MIT license, that description is checkable against real code rather than taken on faith. The part that trips up non-technical readers is the delivery mechanism. Software you’re used to lives on your device or in a browser tab. Hermes needs somewhere to live that doesn’t switch off when you do, which means, for anyone actually using it as an assistant rather than a toy, a small always-on server rather than your laptop. That single fact is what most of this guide is actually about. Does It Actually Run While You Sleep? Yes, with one condition attached: it only runs while you sleep if you deployed it to a VPS, not the desktop app. That distinction is the entire point of Step 2 below, so it’s worth being blunt about it now. The desktop version is tied to a window on your machine. Close the laptop, and every scheduled job, every open task, everything, stops the instant the window does. The VPS version lives on a computer that isn’t yours, in a data centre that isn’t going to sleep, travel, or run out of battery. Once it’s deployed this way, it keeps running for as long as three things stay true: the server is paid and online, your internet-facing dashboard or messaging connection is configured, and your model provider account has funds or an active subscription. That third one is the quiet failure mode worth knowing about early. If a prepaid model balance runs out or a card on file fails, the agent doesn’t crash dramatically, it just stops answering, and you’ll only notice when you check in and find silence instead of a response. Checking your provider’s balance occasionally is a cheap habit worth building once you’re relying on this for real. Six Words You’ll See on Every Screen You don’t need to understand the engineering behind these, but the setup screens use them constantly, so it’s worth ten seconds on each before you start clicking. VPS. Virtual private server. A slice of a computer, sitting in a data centre, that someone else keeps powered on and connected to the internet. You rent it monthly instead of buying a physical machine. Container. A sealed, self-contained box the agent runs inside on that server. Everything the agent touches stays inside the box unless you explicitly let it out, which is what keeps a mistake from reaching anything else on the machine. API key. A password-like string that lets Hermes talk to your AI model provider on your behalf. Treat it like a password, because functionally it is one, and it goes into one specific place in the dashboard, never into a chat message. Tools. Individual actions the agent can take: search the web, read a file, browse a page. One tool, one action. Skills. Saved procedures for combining tools into a finished job. The project’s own documentation describes this as the agent’s procedural memory: when it works out a non-trivial workflow, it can save the approach as a skill and reuse it later without relearning it. Memory. Small, durable facts the agent keeps about you and your setup, stored separately from skills because facts and procedures serve different purposes: memory is what it knows, skills are what it knows how to do. Two Ways to Install This There isn’t one correct way to get Hermes running. There are two, and they solve different problems. Neither is more legitimate than the other, they just answer different questions about what you actually want from it. Option A: Install It on Your Own Computer the Official Way This is the quickest way to get Hermes up and running, and it’s the installation method recommended by the developers. You have two options. If you prefer a simple setup, download the desktop installer from the official Hermes website, run it, and you’re done. No command line is required. If you’re comfortable using a terminal, you can install Hermes with a single command. Mac, Linux, and WSL2 users can run the installation command below, also available here https://hermes-agent.nousresearch.com/ , while Windows users can run the equivalent PowerShell version listed on that same page. curl -fsSL https://hermes-agent.nousresearch.com/install.sh | bash Once the installation is complete, run to choose which AI model Hermes should use. Alternatively, run hermes model if you’d rather use a single subscription instead of setting up separate API keys for different AI providers. After that, type hermes setup --portal to launch it and start chatting. hermes That’s genuinely it. What you’ll notice: this version only runs while your computer is on. Close the laptop, it stops. If you’re trying Hermes out, or you’re fine using it the way you’d use any other app on your machine, that’s not a limitation, it’s just how it works, and you’re already done. The rest of this guide won’t apply to you yet. Option B: No terminal option Run It on a Server Instead, So It Never Turns Off This is the option if you want Hermes reachable at 2am, running a scheduled task while you sleep, or messaging you from your phone without your laptop being involved at all. It means putting Hermes on a computer that isn’t yours, one that’s never switched off. That’s a different piece of infrastructure than installing an app on your own machine, which is why it takes more setup, and it’s what the rest of this guide walks through, using a one-click deployment through a hosting provider so you never have to touch a terminal to get there. If that’s what you’re after, keep reading. Everything from here on is Option B, start to finish. What You’re Trusting It With Hermes can run real commands, and the protection comes from the operating-system-level box those commands run inside. This is why agents that bypass containment or run outside a container represent a real risk. Learn more about what happens when agents confirm things that were never actually agreed to https://industrycontents.com/ai-agent-journey-hallucinations/ . In practice, the one-click Docker deployment in this guide gives you by default the container isolation mentioned above. Before running anything the system flags as risky, Hermes checks it against a list of known dangerous command patterns and, on the default setting, pauses and asks you to approve or deny it. That approval system has three modes: manual, where every flagged command waits for your yes or no; a smart mode, where a secondary model assesses risk and only escalates genuinely uncertain cases to you; and off, which skips the checks entirely and should not be your first setting. Container-based deployment, which is what you’re doing here, reduces the damage a mistake can cause by containing it to a box you can wipe and rebuild, but it isn’t a substitute for the approval system, and it’s worth keeping both in place rather than treating containment as a reason to relax command approval. None of this should put you off. It’s the same reasoning behind any software that can act on your behalf. The one practical rule that follows directly from this is that never bypass the container by running Hermes as a bare process on your own everyday machine, and never disable the approval system just to get past a prompt faster. Learn more about what can go wrong and how to avoid it what-can-go-wrong . What Kind of Work Can It Actually Do? The setup in this guide gets you a general-purpose agent, not a sales agent or a finance agent. The steps are identical no matter which part of the business you have in mind. Nothing in Part 1 is sales-specific or developer-specific. What differs, later, is which skills you teach it and which tools you connect it to. Still, it’s good to know what’s realistic at each stage before you decide if this is worth an afternoon. Keep in mind that agents can route customers toward paths that bypass your revenue model https://industrycontents.com/agentic-ai-business-risk/ without meaning to, so understand the implications of what you’re building. | Function | Possible right after Part 1 | Needs Part 2 or later | |---|---|---| | Sales | Draft a follow-up email or call script when you ask, in conversation | Draft follow-ups on a schedule, or pull context from a connected CRM automatically | | Marketing | Brainstorm campaign angles, summarise research you paste in | Check competitor pages daily and message you what changed, unprompted | | Finance | Answer general questions, run calculations in conversation | Log expenses from a forwarded receipt, produce a recurring weekly summary | | Social media | Draft post copy on request | Post on a schedule, monitor mentions, let you approve drafts from your phone | | General ops and admin | Answer questions with real memory of your business context | Morning briefings, recurring reminders, breaking a task into subtasks automatically | Part 1 gets you a capable conversation partner that remembers your business. Everything in the right-hand column requires either a scheduled job, a messaging channel, or a connection to an outside tool. If you’re worried about how agents perform against financial data, read about how token costs can silently exceed your AI budget https://industrycontents.com/ai-token-costs-budget-overrun/ . Who This Guide Is For You run a business, or you run point on operations for one, and you’ve heard “AI agent” enough times to be curious and skeptical in roughly equal measure. You don’t write code and have no plans to start. You want to know, concretely, what setting one of these up actually involves before you hand it a real task or pay someone else to do it for you. If you’re already comfortable in a terminal, the official quickstart at hermes-agent.nousresearch.com will get you running faster than this guide will. This is for everyone else. What You Need Before You Start - An account with any AI model provider like OpenAI, Anthropic, or an aggregator such as Nous Portal or OpenRouter - 30 to 45 minutes without interruption for the first setup - A password manager, or at minimum a private note, to store the admin credentials you’re about to create - Your credit card for a small monthly server charge Upfront Costs Two separate bills make up the real cost, and conflating them is the most common source of a surprise invoice. One that is fixed is the server that keeps the agent running. The other varies based on the AI model it thinks with, billed by usage, which rises and falls with how much you actually ask of it. This is similar to the token cost overruns that caught Uber’s engineers by surprise https://industrycontents.com/ai-token-costs-budget-overrun/ . | Item | Typical cost | Notes | |---|---|---| | VPS KVM 2: 2 cores, 8GB RAM, 100GB NVMe storage | Roughly $6.49–$9.99/month on a 12–24 month promotional term | Renewal rates run noticeably higher than the intro rate; check the live price at checkout before committing to a term length | | AI model usage, light daily conversation only | Low, typically single-digit dollars monthly on a moderately priced model | Scales with message volume, not with time the server is on | | AI model usage, with frequent scheduled background jobs | Can exceed the server cost | Each cron job and each memory-compression pass consumes tokens even when you never open the chat window yourself | | The Hermes software itself | $0 | Open source, MIT licensed, no subscription for the agent framework | A light setup, checked once or twice a day for simple questions, sits at the low end of that range. An agent running several scheduled jobs a day against an expensive flagship model will spend more on thinking than on hosting. Every setup lands somewhere different on that spectrum, and any source that hands you one confident flat figure is offering a guess dressed up as a fact. Step 1: Choose Your Model Provider Hermes has no intelligence of its own. It’s the body; the model you connect is the brain, and you can swap that brain later without rebuilding anything else. Three starting options cover almost everyone. An existing subscription with OpenAI or Anthropic. If you already pay for ChatGPT or Claude, connecting that account is the path of least friction, and you’ll generate an API key from inside that provider’s own developer settings. Nous Portal. The official aggregator from the team behind Hermes. One subscription covers model access, web search, image generation and text-to-speech, so you’re not collecting five separate keys from five separate companies. OpenRouter. An independent aggregator that routes a single key to hundreds of models from different labs, useful if you want to compare models without opening an account with each one individually. Whichever you pick, what you need before Step 3 is an API key, or in Nous Portal’s case a login you can authorise through, sitting somewhere you can copy it from in the next few minutes. Faster path: if you already use ChatGPT or Claude day to day, note that a ChatGPT Plus or Claude Pro subscription doesn’t automatically include API access – that’s usually billed separately through each provider’s developer platform. If you already have an API key from either, that’s the fast path. If not, budget a few minutes to generate one from OpenAI’s or Anthropic’s developer console before you continue. If this happens: the API key you generate looks like a long string starting with a provider-specific prefix. Copy it once, immediately, since most providers only show it in full one time and won’t display it again afterward. If you lose it before pasting it anywhere, the fix is simply to generate a new one, not to panic. Step 2: Rent Your Server This is the concept so read the first paragraph even if you’re impatient to get to the buttons. A VPS is a computer that lives in someone else’s data centre and stays switched on. You’re not buying hardware. You’re renting a slice of a machine that a hosting company maintains, powers, cools and keeps connected to the internet while you sleep, travel, or simply close your laptop for the weekend. Hostinger currently runs one-click Hermes deployments inside its VPS product, including both the plain Agent template and the fuller Workspace template, which is the path this guide follows because it removes the manual Docker configuration step entirely. Any Docker-capable VPS provider can technically run Hermes; Hostinger is used here specifically because the catalog does the container setup for you. To understand what you’re renting and why, consider reading about how agentic infrastructure is reshaping the business stack https://industrycontents.com/agentic-checkout-adoption-gap/ . Go to the VPS section of the hosting dashboard and select the KVM 2 plan: 2 CPU cores, 8GB RAM, 100GB NVMe storage. That specification is the one independent setup walkthroughs of this exact deployment consistently land on as sufficient for a single agent running a handful of background jobs; a smaller plan will work for pure conversation but tends to strain once you add scheduled tasks in Part 2. Choose a billing term: shorter terms cost more per month but commit you to less, and 12 months is a reasonable middle ground for a first server rather than the steepest discount tier. Pick a data centre location physically close to you or to whoever will be messaging the agent most, since that’s what affects response speed, not the plan tier. Complete checkout. Faster path: don’t overthink the data centre location for a first server. Pick the one closest to your own city and move on; you can migrate later if it turns out most of the traffic is actually coming from somewhere else, but that’s a rare enough case not to plan around it upfront. If this happens: checkout pages for VPS plans often default to the longest, cheapest-per-month term. If you’d rather test the setup before committing to two years, actively look for a 1 or 3-month option before finalising; it costs more per month but costs far less if you decide this isn’t for you after week one. Step 3: Deploy the Hermes Container Once the server is provisioned, you’ll land in its management dashboard. From here the path is: Docker Manager, then Projects, then Compose, then the one-click deployment catalog. Type Hermes into the catalog search. You’ll see more than one result in the catalog. Two of the three are worth a closer look before you click anything. Hermes Agent is the bare core: a working agent with no visual dashboard, just a chat-only screen. Hermes Workspace is a fuller package built on top of it: the same agent, plus a proper dashboard with chat, persistent memory, a skills marketplace, and a built-in terminal, all bundled into one deploy. Deploying Workspace sets up everything it needs on its own; you don’t need to separately deploy Agent first. The third listing, Hermes Web UI , is the one actually worth skipping: it’s a thin front-end that expects a Hermes Agent already running underneath it, and deploying it alone gets you an empty shell with nothing behind it. For this guide, deploy Hermes Workspace . It takes the same one click as plain Agent and gives you a real dashboard instead of a bare chat window, which matters more the less comfortable you are typing commands. Worth knowing honestly, though: Workspace runs more moving parts under the hood, a gateway, a dashboard, and the web interface itself, each its own running service, which means more that can misconfigure compared to the stripped-down Agent option. If you’d rather have the simplest, least-can-go-wrong setup and don’t mind a plainer interface, Agent alone is the safer pick for exactly that reason. Either way, the fields in the next step are the same. If this happens: seeing three near-identical listings and not knowing which to pick is a common point of confusion at this step. The one to actually avoid is Hermes Web UI, since it depends on a separately running Agent it won’t set up for you. Between Agent and Workspace, neither is a wrong choice, just a different tradeoff between simplicity and features. A configuration window opens with a short set of fields: Workspace password. This logs you into the dashboard from here on, not into the agent’s chat itself. Save it in a password manager or a private note immediately; there’s no simpler recovery path than just having written it down correctly the first time. API key field. You can paste the key from Step 1 here now, or leave it blank and add it inside the dashboard after deployment. Leaving it blank and adding it through the dedicated keys tab afterward is the better habit, since that tab is the one place a credential should ever live, and building that habit from day one avoids it becoming a shortcut you take with more sensitive keys later. Faster path: leave the API key field blank here. Adding it in the dashboard in Step 4 takes ten extra seconds and means a typo in this window doesn’t force you to tear down and redeploy the whole container; a typo in the dashboard is a one-field fix. Click deploy. The container builds and starts, typically inside a minute or two. You’ll know it worked when the project shows a running status in the Docker Manager list rather than an error or a stalled build indicator. If it’s still building after several minutes with no progress, refresh the page before assuming something failed; container pulls can be slow on a first deploy depending on server location. Step 4: Connect Your Model From the Docker Manager project row, click Open. This launches the Hermes dashboard in your browser. Log in with the workspace password you just set. Go to Models in the left-hand navigation. If you left the API key blank in Step 3, that field is where it belongs now, pasted into the Keys tab specifically, never typed into a chat message to the agent itself. Once the key is connected, set your main model: the one handling everyday conversation. A balanced, moderately priced model is the sensible default here rather than the newest or most expensive option, since you can change this in two clicks later and the first choice isn’t permanent. Leave the auxiliary tasks setting on automatic. This governs background jobs like compressing long conversations or summarising a document you’ve shared, and the automatic setting routes that housekeeping to a cheaper model so your main model’s cost stays predictable and isn’t quietly inflated by work you never see happening. You’ll know this step worked when the Models screen shows your chosen model as active rather than a connection error. A connection error at this point almost always means the key was mistyped or pasted with a stray space; re-copy it directly from the provider’s dashboard rather than retyping it by hand. If this happens: an empty model dropdown, with nothing to select at all, means the provider connection in the Keys tab didn’t go through, not that something is broken. Go back to Keys, confirm the key shows as connected, and the dropdown will populate. This is a sequencing issue almost every time, not a bug. Step 5: Send Your First Message Open a new chat session from the dashboard and type something as plain as “hi.” A reply confirms the entire chain is working end to end: server running, container healthy, model connected, response returned. This is also the natural point to tell it something true and specific about your business in a couple of sentences, since from here it starts keeping notes, and that persistent memory is the entire difference between this and a chat window you have to re-explain yourself to every single time. If nothing comes back, the fault is almost always in Step 4, not here: recheck that the model shows as connected before assuming the deployment itself failed. That’s the finish line for Part 1. You now have an agent that exists independently of your laptop, answers when you ask it something, and remembers a little about who you are. It can’t yet reach you outside this dashboard, run anything on a schedule, or do much with real tools like your inbox or calendar. That’s Part 2. What Can Go Wrong A handful of mistakes account for most of the friction at this stage, drawn from patterns that show up consistently across independent setup walkthroughs of this same deployment path plus the project’s own documented security guidance. Turning off command approval to move faster. The approval system exists specifically to catch destructive commands before they run. Disabling it removes a real check, not a nag screen, and it’s not the setting to touch just because a prompt is slowing you down. Running Hermes outside its container “just this once.” The container is your main layer of protection, containing a mistaken command to a box you can wipe and rebuild rather than letting it reach the host machine directly. It’s a strong layer, not an absolute one – mounted volumes, network access, and misconfiguration can all still create exposure – which is exactly why it’s meant to work alongside the approval system, not replace it. Pasting an API key into the chat window. Keys belong in the dashboard’s dedicated keys tab, written to the server’s environment file, never typed into a conversation with the agent, since chat content is exactly the kind of thing that can end up logged, summarised, or shared with a tool the agent calls. Undersizing the server, then adding cron jobs in Part 2 and wondering why things slow down. KVM 2 is sized for an agent with a few background jobs running; a cheaper, smaller plan will feel fine for pure conversation and strain once automation is layered on. Choosing the cheapest possible model to save a few dollars, then being disappointed by shallow answers. The main model is what determines conversation quality every single day; it’s not the line item to cut first. Confusing the desktop app with a server deployment. Hermes also ships as a downloadable desktop app, but it only runs while that window is open on your machine. Close the laptop and every scheduled job stops. The VPS path in this guide exists specifically to avoid that limitation, so don’t mix the two expecting server-like behaviour from the desktop version. Losing the workspace password. There’s no shortcut back into the dashboard without it. Write it down correctly the first time. What’s Next Right now your agent can talk and it can remember a little. It can’t yet reach you outside this dashboard, do anything on a schedule without being asked in the moment, or touch real tools like your email or calendar with any depth. Part 2 covers, specifically: connecting it to WhatsApp or Telegram so you can message it from your phone the way you’d message a colleague; setting up its first cron job, a task that runs on a schedule with no one asking; the memory and skill settings that decide what it keeps and what it’s allowed to do unsupervised; and a worked example of teaching it one real recurring task from one of the functions in the table above, start to finish. FAQ Is Hermes Agent free? The software is free and open source under an MIT license. You still pay separately for the server that runs it and for usage of the AI model it connects to, both of which are ongoing costs outside the project itself. Do I need to know how to code to set this up? Not with the deployment path in this guide. The official quickstart uses a single terminal command, but a one-click Docker deployment through a hosting provider’s dashboard avoids the terminal entirely, at the cost of slightly less control than the command-line route offers. What happens if I close my laptop? Nothing, provided you deployed to a VPS rather than running the desktop app. The desktop app only runs while it’s open on your machine; a VPS deployment stays online independently of any device you personally own or carry. Which AI model should I use with Hermes? There’s no single correct answer. A balanced, moderately priced model is a sensible starting point for daily conversation, and Hermes lets you switch models in a couple of clicks later without losing your setup, memory, or skills, so the first choice is a starting point rather than a commitment. Is it safe to run my own AI agent server? Reasonably, if you follow the basics: keep the approval system on rather than disabling it, keep credentials in the dashboard’s keys tab rather than in chat, and rely on the container as your main containment layer, not an absolute guarantee – a command the agent shouldn’t have run generally stays confined to a box you can wipe and rebuild, though misconfiguration, mounted volumes, or exposed network access can weaken that isolation, which is why the approval system stays on rather than being treated as optional. What’s the difference between memory and skills? Memory holds small, durable facts about you and your setup that should always be available. Skills hold longer procedures for finishing a specific kind of task, saved once the agent has worked out how to do it, and loaded only when a job actually calls for them. The project describes memory as what the agent knows and skills as what it knows how to do. Can I move to a bigger server later if I need more power? Yes. Hosting providers generally let you upgrade a VPS plan without a full reinstall, and Hermes itself doesn̻t require a specific plan tier to function, only enough headroom for how many background jobs you’re running at once. Will it keep working overnight or while I’m away? Yes, as long as it’s deployed to a VPS rather than the desktop app, and as long as the server is paid up and your model provider account has funds or an active subscription. The one silent failure mode is a depleted model balance: the agent doesn’t crash, it just stops replying, so check your provider’s balance occasionally once you’re relying on it. Do the setup steps change if I want it for sales versus marketing versus finance? No. Every business function starts from the identical deployment covered in this guide. What differs is what you build on top afterward, which skills you teach it and which tools or platforms you connect, and that starts in Part 2, not here. What’s the actual difference between Hermes Agent and a tool like ChatGPT? ChatGPT is a conversation you initiate. Hermes is infrastructure that stays running, holds tools and scheduled jobs, and reaches you across whichever messaging platforms you connect it to, using whichever underlying model you choose, including OpenAI’s own models if you want them. It’s a different category of product built on top of the same kind of model, not a competing chatbot.