If you have heard about open-source AI agents and it sounds like something for programmers, this article is for you. In Mallorca, more and more medium-sized businesses (a family hotel on Playa de Palma, a clinic with three consulting rooms, a charter company in Port d'Andratx) are asking the same question: what exactly is an AI agent, how does it differ from ChatGPT, and why would anyone want to install one on their own server. Here we explain the two frameworks most widely used in 2026, Hermes Agent and OpenClaw, without any code.
What an AI agent is, in one sentence
An AI chat answers questions. An AI agent does work: it reads your email, checks your booking system, prepares a document, sends it to you on Telegram for approval and, when you say yes, sends it to the customer. The difference is that the agent has access to your tools and can act on its own within the limits you set for it.
An agent framework is the base on which that agent is built. It is like the chassis of a vehicle: the engine (the AI model) is chosen separately and the bodywork (what the agent knows how to do in your business) is fitted on top.
OpenClaw: the most widespread
OpenClaw, created by Peter Steinberger, is in 2026 the most widely used open-source agent framework. Its strength is the ecosystem: there is a huge catalogue of ready-made "skills" and it connects to almost any messaging channel (WhatsApp, Telegram, Slack, Discord and email, among others).
For a business, that means:
- Many pieces have already been solved by others. If you need the agent to read a calendar, check a spreadsheet or post on a social network, there is probably already a skill that does it.
- It is easy to put it where your team already is. If your reception staff live on WhatsApp, the agent lives on WhatsApp.
- There is a large community behind it, with plentiful documentation and examples.
The cost of that breadth is that you have to choose carefully which skills get installed and review them, because not all of them are of the same quality or take the same care with data.
Hermes Agent: the one that learns how you work
Hermes Agent, from Nous Research, was first released in February 2026. Its approach is different: instead of downloading skills from a catalogue, the agent generates its own from what it does with you. It is built around a learning loop: it carries out a task, observes what worked, saves the procedure and does it better next time.
In practice, for a business in Mallorca:
- At the start, you explain how you prepare a charter quote or how you reply to a review. After several repetitions, it has its own written procedure and repeats it consistently.
- It adapts to processes that do not exist in any catalogue, which is the usual situation in a family business with twenty years of its own habits.
- It keeps a memory of what it has learned between sessions, so the knowledge is not lost when the person supervising it changes (relevant where staff are seasonal).
In return, it requires a more careful initial period of guidance, because what it learns depends on what it is taught.
Four terms you will hear, explained
Skill. A specific ability of the agent, written as instructions and small programs. "Check availability in the hotel's PMS" is a skill. "Draft the reply to a review in the house tone" is another. A useful agent usually has between five and twenty.
Cron job. A scheduled task that runs on its own at a fixed time. Every day at 07:30, the agent gathers the overnight bookings and sends you a summary. Every Monday it reviews quotes that have had no reply. The idea is simply "this happens by itself, at this time".
Telegram topic. Inside a Telegram group you can open separate threads by subject. A well-organised agent has one topic for "Approvals", another for "Daily report", another for "Incidents". That way the owner sees what matters to them without everything being mixed into a single chat.
Model provider. The company whose AI does the reasoning: Anthropic (Claude), OpenAI (GPT), Google (Gemini) or xAI (Grok). Both Hermes and OpenClaw let you switch provider without rebuilding the agent. That protects your investment: if a model goes up in price or down in quality, you swap the engine and everything else stays the same.
On your own server, with your own tools
Both frameworks are installed on a server you control, whether one of your own or one we operate in the European Union. Customer data, bookings and emails are not sent out to a third-party platform for storage; the AI model receives only what is needed for each task.
We always build on the tools your business already uses: the PMS or channel manager, the clinic's appointment software, the estate agency's CRM, the quotes spreadsheet, WhatsApp Business. The agent plugs into those; it does not force you to change software or migrate data.
When each one fits
Some guidance based on what we see in businesses on the island:
- OpenClaw fits when most of the work is connecting known channels and systems: receiving guest messages through several channels, answering frequently asked questions, moving information between standard programs. Its catalogue makes for a quick launch.
- Hermes Agent fits when the value lies in a process of your own, repetitive and done by nobody else in quite the same way: preparing charter contracts with the house clauses, qualifying an estate agency's leads against very specific criteria, generating the management report in the exact format the manager wants.
- Many businesses end up with one for the front (messaging) and another for the back office (processes and reporting). They are not mutually exclusive.
In either case, the agent includes an approval loop: nothing is sent to a customer or changed in a system without a person having seen it, until the business decides to relax that control for specific tasks.
Frequently asked questions
Do I need an IT person on staff to have an AI agent?
Not to use it. You do need someone to build it, watch over it and correct it when something changes in your tools; that is the service we offer as a managed agent. Your team interacts with it on Telegram or WhatsApp, just as they would with a colleague.
What happens to my data if I use Claude or GPT as the model?
The model processes the text of each task and returns a result; under these providers' business terms it is not used for training and your data is not retained. The agent's memory, its skills and its history are stored on your server or on ours within the EU. We go into detail in the article on GDPR and automation.
Is it expensive to maintain your own server?
A server adequate for a business agent costs a few tens of euros a month. The real expense is in the use of the AI model (charged by volume of text processed) and in the human work of building and maintaining the agent. That is why it makes sense to start with two or three high-value tasks and expand later.
If you want to know which tasks in your business deserve an agent and which of the two frameworks fits, we always start with an assessment of your current processes and tools. Request an assessment
We implement it for you
We work at the leading edge of artificial intelligence and agents: we build on the most advanced agent platforms of the moment and we do it at your business, in Mallorca, with published prices and a person from your team approving anything that goes out. We start with an initial assessment and a written plan.