Last verified: 2026-08-10 · Primary keyword: Hermes Agent setup · Volatile facts: model names, pricing, and version numbers change frequently — re-check monthly.
What is Hermes Agent and why does it matter?
Hermes Agent is an open-source, self-improving AI agent built by Nous Research and released in February 2026. It is MIT-licensed, runs on a $5 VPS or your laptop, and has accumulated over 222,000 GitHub stars as of August 2026 — making it one of the fastest-growing open-source projects in the AI agent space. Unlike a chatbot that resets to zero every time you close the tab, Hermes carries forward what it learns: your preferences, your project context, and the procedures it has perfected from working with you.
The key differentiator is what Nous Research calls the "learning loop." The agent creates skills from experience, improves them during use, nudges itself to persist knowledge, searches its own past conversations for context, and builds a deepening model of who you are across sessions. This is architecturally distinct from session-based assistants that forget everything between conversations. For anyone building with AI — whether you run a small business, manage content, or write code — this means the work you put in on day one keeps paying you back.
Layer 1: Where does Hermes Agent live?
Hermes Agent runs on a computer. You have three options, and the right choice depends on what you need it to do.
Option 1 — Your own computer. The simplest path. You can download the desktop installer from the official site or run a single curl command on Linux, macOS, or WSL2:
curl -fsSL https://hermes-agent.nousresearch.com/install.sh | bash
The installer takes two to three minutes, requires no prerequisites, and installs everything automatically. If you can copy and paste, you can do this.
Option 2 — A spare machine. An old laptop or desktop you have lying around. Wipe it, install Hermes, leave it on. Your agent now runs all day without touching your main computer. This is useful if you want to sandbox the agent — it can only control what you give it.
Option 3 — A VPS (virtual private server). A small computer in the cloud, available for around $5/month from providers like Hetzner, DigitalOcean, or Fly.io. Your agent stays on 24/7, even when you shut your laptop. The advantage is separation from your personal files — the agent controls only what is on that server. Railway also offers a one-click deployment template for containerized Hermes hosting, with idle memory under 512MB.
Once installed, you connect Hermes to a messaging app. The gateway supports Telegram, Discord, Slack, WhatsApp, Signal, Email, iMessage, and 20+ other platforms from a single process. This is where Hermes stops being a terminal tool and starts being something you can text from your phone while walking the dog.
A quick note on the security question that stops most people before they start: "What if the AI agent goes rogue?" If you run Hermes on a VPS, it can only access what is on that server — not your personal files. You also control permissions inside Hermes. It can read files but not delete them, or draft emails but never send them without your approval. The trust question is a settings menu, not a leap of faith. The agent executes commands under an approval system you configure per tool. For a deeper look at this approach, see our guide on building a local agentic OS that runs a private AI agent on your laptop.
Layer 2: Which model brain should you give Hermes?
Hermes is a body. The AI model is the brain. Hermes lets you plug in almost any model you want — cloud or local — and switching takes seconds. The brain you choose for the job changes everything.
Cloud models
If you already use ChatGPT, you can connect your OpenAI API key straight into Hermes. OpenRouter, Nous Portal, and any OpenAI-compatible endpoint also work. Run hermes model to launch the interactive setup wizard and pick a provider.
One important note: Anthropic restricted third-party apps from using Claude subscriptions, so a Claude Pro subscription will not plug into Hermes the same way an OpenAI key does. You need an Anthropic API key or an OpenRouter key that routes to Claude.
For long autonomous tasks, you want a brain that doesn't quit halfway through the race. Alibaba's Qwen3.8-Max, released August 2026, is a 2.4-trillion-parameter mixture-of-experts model with a 1-million-token context window, costing roughly $2 per million input tokens and $6 per million output tokens via QwenCloud. It was built for long-horizon agent work — in one autonomous coding demonstration, it ran for over 10 days, going from an empty folder to a finished working build without human intervention. For more on that model's capabilities, see our analysis of Qwen 3.8 Max as an autonomous agent model.
Local models
A local model runs on your own machine. Nothing leaves your computer. It is completely private, works offline, and costs nothing per token. The trade-off is that local models are weaker at complex, multi-step reasoning.
The most interesting local model for Hermes right now is Liquid AI's LFM2.5-2.6B — a 2.6-billion-parameter dense model with a 128K token context window and native tool calling. It was trained inside real agent harnesses (including Hermes Agent itself), not on synthetic traces alone. It runs on edge devices, reaches 30 tokens/second on phones and 220 tokens/second on Apple M-series chips, and is fast enough for everyday agentic tasks: calling tools, creating skills, searching the web — all on your machine. You can switch your Wi-Fi off, sit on a plane, and your agent still works.
The layered brain (recommended)
The most effective setup mixes both. Use a frontier cloud model for the important thinking and hand the grunt work to a local model. The boring, repetitive, high-volume stuff goes to the small model fast and for free. The hard decisions go to the big one. It is like a company where the senior person does the strategy and the junior person does the busywork.
Hermes configures itself if you ask it to — you can literally tell it, "Use the strong model for important work and the local model for background jobs," and it writes the config. You can also switch between models in the desktop app's dropdown. The key advantage: when a new model drops next month — and it will — you swap the brain and keep the body, the memory, and every skill you have built. Nothing gets lost.
For more on running a fully local agent stack, see our guide to the BTL-4 and Macaw free local AI agent stack.
Layer 3: How does Hermes memory work?
This is the layer that makes Hermes fundamentally different from ChatGPT, Claude, or any session-based assistant. Memory has three levels.
Level 1 — Durable fact files. Hermes keeps a small set of files under ~/.hermes/. One file (MEMORY.md) holds notes the agent keeps for itself: what is set up, what is in progress, lessons it has learned. Another (USER.md) holds facts about you: your name, your business, how you like your replies. These files load at the start of every session, so your agent knows who it is already working for before the first message. The files are deliberately kept small — because if everything is memory, nothing is memory. Hermes only keeps the facts that matter, so it stays fast instead of drowning in old chat logs.
Level 2 — Session search. Hermes quietly saves your past conversations in a SQLite database using FTS5 full-text search. It does not cram them into every chat. But if you ask, "What did we decide about that pricing idea last week?" it searches its own history and pulls the answer. This gives you long-term recall without inflating the context window.
Level 3 — Automated memory updates. Every few messages, Hermes quietly asks itself: "Did I just learn something worth keeping?" If yes, it saves it. You do not have to manage any of this. You just talk, and it gets to know you. There is also a security check built into that loop: before anything gets written to memory, Hermes scans it for prompt injections and hidden characters, so nothing sneaky gets planted in the agent's brain.
You can also connect Hermes to Obsidian, a free note-taking app, and turn it into a second brain. Every piece of research, every project, every output gets saved as readable notes that you can open yourself — and your agents can read those notes back later and build on them. This gets powerful when you run multiple agents: the vault becomes the shared memory for every agent in the system. For a detailed walkthrough of this pattern, see our guide to Claude Obsidian 2.0 and persistent AI memory.
Layer 4: How do Hermes skills work?
If memory is what your agent knows, skills are what your agent can do. This is the feature that made Hermes famous.
A skill is a step-by-step instruction sheet for one task, stored as a markdown file under ~/.hermes/skills/. Hermes comes with 40+ built-in skills — managing notes, doing research, creating files, and more. But the real power is that it builds new skills from working with you.
Here is how it plays out:
- You ask Hermes to research your competitors. It does an okay job.
- You give it feedback: "Check the newest post. Compare them to mine. Keep it short." It does it better.
- You do this a few times. Hermes notices the pattern and says, "Should I save this as a skill?" You say yes.
- The entire process — with all your feedback baked in — is saved forever as a
SKILL.mdfile. - Next time you say "Run my competitor check," it does the whole thing your way, first try.
There is also a faster way to teach it: the /learn command. You type /learn inside Hermes and hand it a guide — a tutorial, documentation page, or article. Hermes reads it, works out the process, and writes itself a skill file on the spot. Every guide on the internet just became a training manual for your agent. If you found a great article about a new tool this morning, your agent knows how to use it by lunch.
Hermes also keeps its own skill library clean. A feature called the Curator watches for skills you never use. Unused for a month? It gets flagged. Unused for three months? It gets archived (not deleted — always restorable). Your agent tidies its own desk. You can pin critical skills to protect them from ever being archived.
The compound effect is real: with a normal chatbot, you correct it forever and get the same mistakes every week. With Hermes, you correct it once and the correction sticks. Every week that passes, your agent knows more recipes. For more on building agent skills that actually compound, see how to build Claude AI marketing skills as reusable workflows.
Layer 5: How do you schedule work with Hermes cron jobs?
This is where Hermes stops being a chat and starts becoming an employee. A chatbot writes an answer and waits for you to show up. Hermes shows up with the work already done.
The feature is called cron jobs — just a nerdy name for scheduled tasks. You tell Hermes in plain English, "Do this task every morning at 8," and it does it forever without being asked twice.
# Create a daily morning brief
hermes cron create "0 8 * * 1-5" "Check what matters to my business today and send me a summary" --deliver telegram
# A weekly competitor scan
hermes cron create "0 9 * * 1" "Scan my competitors' websites. Compare this week to last week. Report only what's changed — what they published, what they launched." --deliver local
Two smarter versions of this are worth setting up:
Exception watching. Instead of sending you a report every day, the agent watches quietly and only messages you when something changed that matters — a review appeared, a competitor posted something big, a price dropped. This cuts notification noise to almost zero while keeping you informed.
Research digests. Every 24 hours, the agent finds the latest headlines in your industry, orders them by priority, cites the source, and delivers a clean summary. This runs without you and plugs into your memory system, so it is ready to go when you need it.
One maintenance tip: clean out old jobs once a week. Too many scheduled tasks running in the background will slow your agent down. Pause anything you're not using, and your agent instantly gets faster. For a practical playbook on running daily AI agent operations, see our agent OS daily operations playbook.
Layer 6: How do you run a team of Hermes agents?
This is the advanced level, and where things get genuinely fun. Three concepts build on each other.
Subagents
When Hermes has a big job, it can split the work and spin up helper agents that run at the same time. Say you're researching a new offer for your business. One subagent studies your competitors. Another checks what people are searching for. A third reviews what you have already made. Then the main agent pulls it all together and gives you one answer. What took you a full afternoon now comes back whilst you made lunch.
Give each subagent a narrow job with a clear output. This is also where a local model like LFM2.5 works nicely: your subagents do the grunt work on the small fast local model, while the main agent thinks with the big brain. For a deep dive on structuring this correctly, see our guide to graph engineering for AI agents.
Profiles
A profile is a separate Hermes with its own memory, its own skills, and its own job. You can run a content agent that knows your voice, a research agent that knows your market, and an operations agent that knows your systems. One assistant shouldn't hold everything in one brain — the same way one employee shouldn't run every department.
Create profiles with:
hermes profile create content --clone
hermes profile create research --clone
hermes profile create ops --clone
hermes profile list
Each profile gets its own model. Your content profile can run on the big brain; your background research profile on the local one. For a full walkthrough, see our AI agent OS setup guide for running a 24/7 AI team.
Goal mode
Instead of giving your agent a task, you give it a goal — a finished outcome — and it keeps working step by step until the goal is met. This is not one reply and done. It can take hours if it takes hours. You set the target, walk away, and come back to it finished. This is exactly why the brain choice matters so much: a model built for long autonomous runs (like Qwen3.8-Max) paired with goal mode means you can hand over a genuinely big piece of work in the morning and return to it done.
One tip before you run a goal: ask Hermes to help you write the best possible goal for what you want. It writes its own instructions better than you will.
Mixture of Agents
Mixture of Agents is a feature where Hermes takes your task, sends it to more than one model at the same time, and fuses the answers into one better output. Two brains attack the same problem, then merge the best of both. When the task really matters, this squeezes out quality that no single model gives you alone.
You select a Mixture of Agents preset the same way you select any model:
/model review --provider moa
MoA composes automatically with goal mode, gateway sessions, and the desktop app — it is a normal provider in the model system, not a separate mode.
What MCP connections add
Hermes supports the Model Context Protocol (MCP) — a standard way to plug outside apps into your agent. That could be your email, your calendar, your notes, your image tools, automation platforms, or databases like Supabase. Once connected, Hermes can use them the way an employee uses software on a computer.
The setup lets you control exactly what the agent is allowed to do per app. It can find emails and write drafts but not send or delete. You choose the permissions app by app. For what changed in the MCP spec recently and how to migrate, see our MCP stateless protocol migration guide.
The seven mistakes that make Hermes worse
The fastest way to ruin this is to set it up in the wrong order or with the wrong priorities. These are the seven mistakes that get most people.
| # | Mistake | Why it backfires | What to do instead |
|---|---|---|---|
| 1 | Adding too many tools before you know what you want | You clutter the agent's context and decision space before it has a clear job | Start with the tools your first workflow actually needs |
| 2 | Saving everything to memory | Memory becomes a dumping ground, the signal drowns in noise | Keep memory for the facts that matter — short, durable, reusable |
| 3 | Using a weak model for hard work | A small model failing and retrying five times wastes more than a strong model doing it once | Match the model to the task; use the frontier brain for hard work |
| 4 | Turning every idea into a scheduled task | Your agent gets bogged down running low-value jobs that never end in a decision | Only schedule tasks that end in a decision or an action |
| 5 | Trusting subagent output without checking | AI is still early; subagent summaries can be wrong or fabricated | Check their summaries; don't believe everything the agent tells you |
| 6 | Building specialist profiles before you have one working workflow | You spend time configuring agents that have no real job yet | Get one workflow working first; profiles come later |
| 7 | Treating Hermes like a magic employee | It runs on good instructions, good memory, and good skills — all of which take time | Feed it well and it improves; feed it chaos and it compounds chaos |
What this means for you
If you run a small business, manage content, or write code, the practical takeaway is this: the setup takes an afternoon, and the tasks it takes over stack up for months. You install Hermes, connect one model, wire it to your phone, teach it one skill, and schedule one recurring task. Then you do it again. Each layer compounds.
The gap between you and the people already winning with this is not technical skill — if you can send a text message, you have every skill this requires. The gap is that they started. The person who started three months ago has three months of compound learning that you can't get back.
Your 7-day implementation plan
| Day | Action | Outcome |
|---|---|---|
| 1 | Install Hermes and connect one model | Agent runs, responds in CLI |
| 2 | Connect it to your phone via Telegram or WhatsApp | You can chat from anywhere; run one real task |
| 3 | Take a task you repeat weekly and teach it as a skill (or use /learn on a guide you love) |
One reusable skill saved forever |
| 4 | Connect the one MCP tool that matters most for your work | Agent can read your email, calendar, or database |
| 5 | Create one scheduled task that gives you a decision, not a data dump | A morning brief or exception alert running automatically |
| 6 | Split a research task across subagents, or run your first goal-mode session | Hours of work happen without you |
| 7 | Create your first specialist profile | One useful workflow, then another, then another |
FAQ
Q: Is Hermes Agent really free? A: Yes. Hermes Agent is MIT-licensed and open source — you can run it free forever on your own hardware. The only cost is the model API calls (if you use a cloud model) or your compute (if you run a local model). OpenRouter offers free models at zero cost to get started.
Q: Can I run Hermes Agent fully offline? A: Yes. Install Hermes, pair it with a local model like LFM2.5-2.6B via Ollama or LM Studio, and disconnect the network. Memory, skills, cron, and file tools all work offline. Complex multi-step workflows may degrade with local models, but single-shot chat and coding work fine.
Q: How is Hermes different from ChatGPT or Claude? A: ChatGPT and Claude are session-based — they forget everything when you close the tab. Hermes has persistent memory across restarts, creates reusable skills from experience, runs scheduled tasks without you, and connects to 20+ messaging platforms. It is an agent that runs unattended, not a conversation that ends.
Q: Do I need to be technical to use Hermes? A: No. The install is one curl command (or a desktop installer). Configuration is done by talking to the agent in plain English — "Use the strong model for important work and the local model for background jobs" — and it configures itself. The dashboard provides a plug-and-play UI for channels, MCP, skills, and scheduled tasks.
Q: What is the best model to use with Hermes Agent? A: It depends on the task. For long autonomous runs, a frontier model like Qwen3.8-Max (2.4T params, 1M context, ~$2/$6 per million tokens) is built for sustained agent work. For everyday tasks on your own machine, Liquid AI's LFM2.5-2.6B (2.6B params, 128K context, free) is trained inside agent harnesses. The recommended setup is to mix both: the frontier model for hard thinking, the local model for grunt work.
Q: Can Hermes Agent run on a VPS? A: Yes. A $5/month VPS is one of the best options because the agent stays on 24/7, even when you shut your laptop. Railway offers a one-click deployment template, and Hermes uses under 512MB of memory at idle. Running on a VPS also isolates the agent from your personal files — it can only control what is on that server.
Every claim here is traced to a primary source, dated, and listed under Sources. Research and drafting are AI-assisted; editing, verification and publication are human decisions, and a person is accountable for what appears on this page. How we work →
Discussion
0 comments