OpenClaw vs Hermes: which AI agent to choose in 2026

Imagine sending a WhatsApp message saying "organize my inbox and let me know what's urgent" and receiving, minutes later, the task done. Not by a person, but by a program running on your computer, reading your emails and acting on its own. That's the promise of personal artificial intelligence agents, and in 2026 two names dominate the conversation.

The question for newcomers is always the same: OpenClaw vs Hermes, which to choose? Both are free, open source (the code is public, and anyone can use, inspect and modify it) and do similar things. But they start from different ideas about what an agent should be.

In this guide you'll understand what each one is, where they differ, how much they really cost, what risks they bring and which fits your profile.

Quick summary: the answer in 30 seconds. OpenClaw and Hermes Agent are free AI agents that run on your machine and talk to you through apps like WhatsApp and Telegram. OpenClaw bets on reach: it connects to more apps and has the largest community. Hermes bets on memory: it learns from use and creates new skills on its own. Both give the agent real access to your computer, so security comes first.

Index

• What are OpenClaw and Hermes?
• OpenClaw vs Hermes: what's the difference?
• How does each remember you and learn on its own?
• Where do they run and through which apps do you talk to them?
• How much does it cost to use OpenClaw or Hermes?
• Are OpenClaw and Hermes safe?
• OpenClaw or Hermes: which to choose?
• Frequently asked questions

What are OpenClaw and Hermes?

OpenClaw and Hermes Agent are open-source personal AI agents. Instead of just answering questions, like a chatbot, they execute tasks on your computer and in your apps: they read files, send messages, browse the web and run commands (orders typed directly to the computer). You talk to them through your phone, as you would with a human assistant.

To understand the difference between a chatbot and an agent, think of a consultant and an intern. The consultant (regular ChatGPT) tells you how to solve the problem. The intern (the agent) takes the task and solves it. If you want the complete explanation of this concept, Kalkazu has already shown what agentic AI is from scratch.

An important detail: neither is a "brain". They are what programmers call a harness (support structure): the body and hands of an AI. The brain is a language model, like Claude, GPT or Gemini, which you choose and connect to the agent. The body gives that brain access to files, apps and the internet.

This matters because the quality of responses depends heavily on the model you use. OpenClaw and Hermes decide what the model can do and how it remembers you. That's where they differ.

What this body knows how to do comes from skills: instruction packages that teach the agent to complete a task. Think of a recipe stored in a drawer, which it only opens when needed.

In one sentence: both are "bodies" for an AI brain you choose, and the difference lies in what that body knows how to do.

OpenClaw: the agent that went viral

OpenClaw was born as a personal project by Austrian programmer Peter Steinberger and changed names twice. It started as Clawdbot, became Moltbot and, according to the official project history, got its definitive name on January 30, 2026. The mascot is a lobster, and the community adopted the creature as a symbol.

The explosion was rapid. The Verge followed the agent becoming an obsession in the tech world at the beginning of the year, with people using the program to clean their email inbox, check in for flights and manage their calendar.

The project page on GitHub, the site where programmers store and share code, has over 390 thousand stars (checked on 09/25/2026). The star works like a "favorite", and few platform projects have reached this high.

Steinberger now works at OpenAI, but the project doesn't belong to the company. Since July 2026, OpenClaw has been maintained by the OpenClaw Foundation, a nonprofit organization in the United States. The official site makes it clear: OpenAI is one of the donors, "OpenClaw is not an OpenAI product" and there is no paid version.

At the end of August came OpenClaw 2.0, the biggest update in the project's history. There were over 16 thousand code changes, submitted by 933 people, with simplified installation and a rebuilt dashboard, the page you open in the browser to control the agent.

Hermes Agent: Nous Research's answer

Hermes Agent is the agent from Nous Research, an AI lab known for the open language models of the Hermes family. The oldest version on the project's releases page is from March 2026. In other words, Hermes arrived when OpenClaw was already a phenomenon.

The slogan sums up the proposal: "the agent that grows with you". On the official page, Nous promises persistent memory and an agent that "learns your projects, generates skills on its own and never forgets how it solved a problem".

Growth was also strong. The Hermes page on GitHub has over 248 thousand stars (checked on 09/25/2026). TechCrunch describes Hermes as OpenClaw's direct competitor, with an important difference: it already comes with ready-made skills and was designed to create others with use, without manual intervention.

OpenClaw vs Hermes: what's the difference?

The main difference lies in each one's bet. OpenClaw bets on reach: more messaging apps, the largest community and a large store of third-party skills. Hermes bets on memory and learning: it creates its own skills after complex tasks and improves with use.

In practice, both do many of the same things. Both are free, run on your computer or on a server (a rented computer, always on), accept various AI models and talk through your phone. The table shows where they diverge (data checked on 09/25/2026).

One confusing point: the programming language. It doesn't change anything for those who will just use it. It matters for those who want to modify the code, because Hermes speaks the most common language of AI research, Python, and OpenClaw speaks the language of the web, TypeScript.

In one sentence: OpenClaw is the Swiss Army knife with more blades; Hermes is the assistant that learns your way.

How does each remember you and learn on its own?

Both have memory between conversations, but Hermes makes memory the center of the product. It stores what it learned, searches old conversations and, after a complex task, writes a new skill on its own to repeat the feat. OpenClaw also remembers and can create skills, but depends more on the community store.

A skill can teach, for example, how to generate a weekly report from your calendar.

OpenClaw's path: the community store

OpenClaw bets on the power of the crowd. Anyone can publish a skill on ClawHub, the project's store, and you install whichever you want. The official site says the agent remembers your preferences, projects and people from one conversation to another, and that the agent can even create its own skills.

The advantage is variety: if someone has already solved your problem, the skill probably exists. The disadvantage is that you're installing code from strangers, and this has already caused problems, as you'll see in the security section.

Hermes's path: the learning loop

Hermes was designed around what Nous calls a closed learning loop: it does, notes what worked and uses the note next time, without needing anyone from outside. According to the official repository, it creates skills from experience, improves these skills during use, searches its own past conversations and builds, session after session, an increasingly detailed profile of who you are.

A concrete example: you ask Hermes to gather the month's invoices into a spreadsheet. It discovers the path, makes mistakes, corrects and finishes. Next time, instead of starting from scratch, it uses the skill it wrote the first time. The skills documentation confirms that skills created by the agent are saved in a folder on your computer, along with those you installed. This way, the agent becomes more useful over time without you needing to configure anything.

In one sentence: OpenClaw learns from the community; Hermes learns from you.

Where do they run and through which apps do you talk to them?

Both run on Mac, Windows and Linux, and also on cloud servers. You talk to them through messaging apps or through a dedicated app. OpenClaw connects with more apps, including iMessage, and has an official mobile app. Hermes runs on more types of servers.

Where the agent lives

A personal agent needs to be on to work. That's why many people don't install it on their laptop, which turns off, but on a VPS (virtual private server): a rented computer in the cloud, on 24 hours, that costs a few dollars a month. It's like leaving the intern in their own room, instead of at your desk.

Hermes highlights this flexibility. The repository says it works both on a $5 per month VPS and on a set of graphics cards (the parts that make AI run fast), and also on services that "hibernate" the agent when it's idle, charging almost nothing. Nous also offers Hermes Desktop, an app for Mac and Windows.

OpenClaw has apps for Mac, Windows and Linux on the official site, and in June got apps for iPhone and Android, according to TechCrunch. The mobile app connects to the agent already running on your computer and allows you to approve its actions remotely.

Where you talk

For users worldwide, WhatsApp support matters most, and both deliver. If you live in the Apple ecosystem and want to talk to the agent through iMessage, OpenClaw has the advantage.

How much does it cost to use OpenClaw or Hermes?

Both programs are free, but using them isn't free. You pay for the brain, the AI model, either by subscription or by use. If you want the agent on all the time, add the server. Hermes also has an optional Nous plan, from $20 to $200 per month.

The agent, alone, doesn't think. Each request you make becomes a query to the language model, and that query has a cost. The more tasks the agent executes on its own, the more it consumes.

The cost in OpenClaw

OpenClaw doesn't sell anything. The official site states there is no enterprise edition or paid version.

You connect the model you prefer, reusing an AI subscription you already have or using an API (application programming interface): the door through which one program uses another and pays per use, like an electricity bill.

For this you create an API key, a kind of password that links billing to your account. The OpenClaw documentation warns that the model company can change prices and limits at any time, even if the agent stays the same. Worth checking before subscribing.

The cost in Hermes

Hermes is also free and accepts any provider (the company that supplies the model), like OpenRouter (a service that gives access to various models with one account), OpenAI or your own server. The difference is that Nous offers a paid shortcut, Nous Portal, which brings together in one place the models and extra tools (web search, image generation, voice).

Nous Portal plans, checked on 09/25/2026 on the Hermes page:

Notice how it works: you buy credits to spend with the agent, not unlimited conversation. Every paid plan gives 10% more in credits than it costs. If the agent works a lot, the credits run out before the end of the month.

An example of monthly bill

A lean scenario running 24 hours: Hermes on a $5 per month VPS, with the Plus plan at $20. The bill comes to $25 per month, as long as the agent doesn't spend more than the plan's credits. In OpenClaw, the similar scenario is a server of the same type plus what you already pay for the model, whether subscription or API use.

If you already have an AI subscription, it may be cheaper to reuse it. Just keep an eye on limits: the OpenClaw documentation reminds, in Claude's case, that use through third-party apps is deducted from subscription limits, not charged separately.

In one sentence: the program is free in both, and the real bill is the AI model you plug into them.

Are OpenClaw and Hermes safe?

Neither is safe by default, because the risk lies in the very idea of an agent. To be useful, it needs to access your emails, files and accounts, and that makes it a target. OpenClaw has already had skills with viruses and installations exposed on the internet. Hermes appeared in a study about agents that installed dangerous code.

The risk that comes with any agent

The biggest danger has a name: prompt injection. An AI model doesn't know how to separate well what is your order from what is text it read somewhere. So someone hides an order in an email or on a page, and the agent obeys.

It's like an intern who reads an anonymous note on the desk saying "transfer $500 to this account" and thinks it came from the boss. Researcher Simon Willison calls the combination that makes this serious the lethal trifecta: access to private data, contact with strangers' content and ability to send things out. OpenClaw and Hermes have all three.

OpenClaw's track record

Fame brought problems early. In February, researchers, including those from VirusTotal, a service that checks if files have viruses, found hundreds of malicious skills on ClawHub, which stole data from those who installed them. The project responded with a partnership with VirusTotal, which began scanning all store skills daily.

The announcement itself was honest: scanning doesn't solve everything. It catches known viruses, but doesn't catch a skill that brings, written in plain text, an order for the agent to do something bad. There were also those who exposed the agent on the internet without a password, leaking conversations and access passwords to services, as shown by VentureBeat.

Experts interviewed by TechCrunch were harsh. One of them summarized the agent as someone "sitting with a bunch of passwords on a machine connected to everything".

Hermes's track record

Hermes has fewer public episodes, partly because it's newer. But it's not free from problems. In August, Ars Technica reported a study about llms.txt, a file that sites publish to guide AIs.

In some of them, the instructions ordered installation of code packages that didn't exist. It's like a manual that indicates an address no one occupied: any scammer can move in there. Researchers occupied some of these addresses to test, and agents like Hermes, Claude and Codex (OpenAI's programming agent) followed the instructions and installed what was there.

As general protection, the Hermes documentation states that every skill installed from the official store goes through a security checker, which looks for data leakage, prompt injection and destructive commands. It doesn't cover, however, the llms.txt case, which doesn't go through the store.

Help from Windows

Microsoft entered the subject in June. At Build 2026, the company's annual conference for developers, it introduced MXC, a system that creates a sandbox (a closed space where the agent works without touching the rest of the computer). On Windows, OpenClaw already works inside this sandbox, and Nous announced that Hermes will adopt it in the Windows app.

Precautions before installing either

1. Run the agent on a separate computer or server, without your personal files.
2. Create new accounts for it (an email just for it), instead of giving access to yours.
3. Install only skills from known sources and read what they do.
4. Enable manual approval for sensitive actions, like paying or deleting.
5. Never leave the agent's dashboard (its control page) open on the internet without a password.

In one sentence: treat the agent like a new intern, with access only to what it needs, until it proves it deserves more.

OpenClaw or Hermes: which to choose?

Choose OpenClaw if you want the largest number of connected apps, the largest community and lots of help content. Choose Hermes if you want an agent that learns from use and gets better on its own, or if you prefer to pay a single plan and not deal with multiple API keys.

If you're a beginner

Neither works just by installing. You need to open the terminal (the screen where you type commands), paste a command and follow a configuration wizard. Desktop apps make it easier, but you'll still deal with API keys and permissions. If you've never opened a terminal, start slowly and on a test computer.

If you already use OpenClaw

You can try Hermes without losing what you've already built. The official repository brings the hermes claw migrate command, which imports from OpenClaw your memories, skills, messaging settings and API keys. There's even a simulation option, which shows what would be copied before copying.

One last observation, an opinion: for most readers, the question isn't which of the two, but whether it's worth having an agent now. If your daily routine doesn't have repetitive tasks on the computer, a regular chatbot solves it, with much less risk. To see the bigger picture, Kalkazu's artificial intelligence category gathers the other guides.

Frequently asked questions

Are OpenClaw and Hermes free?

The programs are free and open source, with MIT license. What costs is the AI model you connect to them, by subscription or by use. Hermes also offers Nous Portal, an optional plan from $20 to $200 per month.

Do I need to know how to program to use them?

You don't need to program, but you need patience with configuration. Installation goes through the terminal or through a desktop app, and you'll choose model, apps and permissions. Those who've never done anything similar should set aside an afternoon and use a test computer.

Do both work in Portuguese?

Yes. The language depends on the connected AI model, not the agent. Models like Claude, GPT and Gemini understand and respond well in Portuguese, so you can talk normally through WhatsApp.

Can I use OpenClaw and Hermes at the same time?

You can, but it rarely pays off. Both would use the same WhatsApp and the same email, and you'd pay for the model twice. The most common path is to test one, and, if you want to switch, use Hermes's migration command to bring memories and skills from OpenClaw.

Do OpenClaw and Hermes run on mobile?

The agent itself runs on a computer or server. OpenClaw has official apps for iPhone and Android that connect to that agent, and you can talk to both through WhatsApp or Telegram. In practice, the phone becomes the remote control.

OpenClaw and Hermes show two bets for the same future, and the best choice is the one that fits your way of working. Start small, on a separate computer, and give the agent more access only when it deserves it.