Hosting agentów AI. Twój agent, Twój serwer, Twój klucz.

1. Your AI agent does not have to live in someone else's cloud

Most popular AI assistants operate as SaaS services, with an API key, a monthly query limit, and conversation data passing through third-party servers. For many business use cases, such as customer support or internal automation that accesses company data, this raises the question of where that data actually ends up.

The Saplo app catalog includes several ready-to-use engines for building your own agents and AI automations. They run on your own Box, under your own domain, and avoid third-party SaaS infrastructure except for the language model the agent connects to.

The flagship app is Hermes Agent (Nous Research), a personal AI agent that connects to messaging platforms such as Telegram, Discord, Slack, and API. Its web dashboard lets you configure the model key and messaging integrations without accessing the server itself. The manifest requires approximately 1.5 GB of RAM.

2. Four different engines, four different approaches to AI

Hermes Agent is a ready-to-use, preconfigured messaging agent. It is ideal if you want an assistant on Telegram or Discord without building anything from scratch, simply by entering your model key and connecting a messenger in the app dashboard.

Flowise is a drag-and-drop visual builder for AI agents and chatbots. It combines LLM models, tools such as API calls and functions, and RAG vector databases into a node graph without requiring code. The manifest requires 2 GB of RAM because the image includes Chromium for nodes that scrape web pages and PDF files.

Although n8n is not a dedicated AI engine, it often acts as the glue between an AI agent and the rest of a company's systems. A webhook from Hermes Agent or Flowise can trigger an n8n workflow that saves the result in a CRM or sends a notification, with all apps running on the same Box.

Windmill is a code-based automation platform supporting Python, TypeScript, Go, and Bash. It is closer to a developer's tool than a visual flow builder such as n8n or Flowise, and is useful when an agent's logic requires a custom script rather than ready-made node blocks.

3. How much capacity does your own AI agent infrastructure need?

Hermes Agent (1.5 GB RAM minimum) and Flowise (2 GB RAM minimum) together exceed the capacity of Box M, so running both on one Box realistically calls for Box L (6 GB RAM), leaving room for the operating system and other apps running at the same time.

Hermes Agent on its own, without other resource-intensive apps, fits on Box M (2 GB RAM). That is enough if you only want a single messaging agent, with no additional visual engines or automation tools alongside it.

The cost of using the language model, billed separately through your LLM provider's API key, is separate from the hosting cost, just as with any other self-hosted AI solution. Saplo hosts the agent engine, not the model the agent talks to.

4. Why this is different from a SaaS subscription for an AI agent

An AI assistant SaaS subscription is usually billed per user or by the number of monthly queries. As usage grows across a company, the monthly cost rises too, often with no clear upper limit.

Your own instance on Saplo has a fixed annual cost, based on the price of the Box, regardless of how many messages the agent handles each month. That makes budgeting easier for companies with irregular or seasonal traffic.

Conversation data and agent configuration remain in the LXC container on your Box, not in the SaaS provider's database. This is important for companies processing customer personal data through an agent, such as in customer support, when assessing GDPR compliance.

5. Who it is for

For a small business that wants a simple Telegram assistant for customer questions, Box M with Hermes Agent is a quick, affordable way to get started without building anything from scratch.

For a technical team building a more complex flow that combines an agent, automation, and CRM integrations, Box L with Hermes Agent, Flowise or Windmill, and n8n together on a single Box gives you full control over the entire AI stack.

If you're testing which AI engine best fits your company's needs, the catalog lets you deploy several engines at once on a single Box and compare them without separate SaaS contracts for each one.

6. How to get started with your own AI agent, step by step

Choose the engine for the specific use case: Hermes Agent for a simple messaging assistant, Flowise for a more complex agent with access to documents (RAG), and Windmill when your logic requires custom code rather than just node blocks.

Install your chosen app from the catalog on the right Box, with Box M as the minimum for Hermes Agent alone or Box L when combining multiple engines. This deploys a Docker container with a ready-made reverse proxy and SSL configuration.

Configure the language model API key in your chosen app's dashboard without accessing the server itself. It is designed so that a nontechnical person at your company can connect a messenger or set a system prompt independently.

Test the agent with a small group of users, such as an internal test channel on Discord, before making it available to customers. This helps you catch prompt configuration or integration issues before real customer interactions.

Connect the agent to your other systems through n8n running on the same Box, if you want its responses to continue to a CRM or database, or trigger other business automations.

7. Data privacy when using AI agents in business

Conversations handled by an AI agent, such as customer questions on Telegram, are stored in a container on your Box, not in an external SaaS provider's database. This gives you tighter control over who can access the data.

The language model your agent talks to is usually an external service, specifically the LLM provider's API, so the content of your queries still reaches that provider. This is worth considering when assessing GDPR compliance for personal data in conversations.

For businesses processing sensitive customer data through an agent, consider language model providers with a clear policy against using queries for further model training, regardless of where the agent engine itself is hosted.

A self-hosted agent engine (Hermes Agent, Flowise) gives you control over logs and conversation history at the infrastructure level, which a fully external SaaS chatbot platform does not always guarantee.

8. Integrating an AI agent with your existing business systems

AI agents rarely operate in full isolation from the rest of your business systems. More often, they need to check product availability in a database, create a ticket in a CRM, or route a more complex customer question to a specialist. This requires connecting the agent to other apps, not just to a language model.

In Saplo, this connection is most often implemented with n8n running on the same Box. A webhook from Hermes Agent or Flowise triggers a workflow that checks the database, saves data to a CRM such as Baserow, or sends a Slack or Discord notification to the right team member.

This architecture, with an agent, automation, and database all on one Box, lets you build relatively sophisticated customer support scenarios without writing custom integration code, using only ready-made engines from the app catalog connected through standard webhooks and APIs.

If your business already uses a CRM outside Saplo, such as a third-party SaaS, the agent and automation can connect through its public API, if the CRM provider offers one. This preserves your existing infrastructure and adds an AI layer without migrating the entire CRM.

Start by testing the integration with a simple scenario, such as having the agent save only the customer's question to the database without automatically answering it. Once you've confirmed that the data reaches the right place, add more advanced agent actions.

As integrations grow, documenting which n8n workflows handle which agents and which systems they connect becomes essential. Without documentation, several interconnected apps on one Box can quickly become a hard-to-understand web of dependencies for anyone joining the project later.

9. The open source ecosystem for AI agents

Hermes Agent, Flowise, Windmill, and similar engines are actively developed open source projects. This means regular releases with new features and integrations for more language models, which you can use in your own instance without waiting for a SaaS provider to make them available.

Development in the AI agent space is exceptionally fast compared with other software categories. Using actively maintained open engines means you are more likely to gain access to new capabilities as soon as they are released, rather than waiting for a closed commercial platform to update.

On Saplo, you can update these engines to newer versions using a recipe from the app catalog. This makes it easy to stay up to date without tracking each tool's GitHub repository separately.

For teams experimenting with different language models, such as switching LLM providers or testing new models, a self-hosted agent engine lets you connect any API key without the restrictions imposed by a closed platform tied to one specific model provider.

The community around these projects, including forums, repositories with examples, and Discord channels for open source projects, is another source of knowledge for solving less common use cases. It provides knowledge independent of Saplo's documentation, which focuses on the infrastructure layer rather than configuring the agent engine itself.

10. Measuring AI agent performance after deployment

After implementing a messaging agent (e.g. Hermes Agent on Telegram), track the percentage of customer inquiries the agent handles on its own and the percentage that need to be handed off to a human. This gives you a concrete measure of whether the agent is genuinely reducing your customer support team's workload.

Regularly reviewing conversations after anonymizing them to protect customer privacy helps identify recurring questions the agent cannot answer well. This signals that you should improve the system prompt or add information to the agent's knowledge base. This cycle of reviewing, refining, and reviewing again is a natural way to improve the agent over time, just like with any other product that interacts directly with your customers.

For agents built in Flowise with access to company documents (RAG), periodically check that the agent's knowledge base is up to date. Outdated documents can lead to outdated and potentially incorrect answers for customers.

Evaluate effectiveness not only by the system's technical performance but also by customer satisfaction with their conversations. Even a technically correct answer can feel too stiff or may not match the customer's communication style.

11. Summary: your own AI infrastructure as an advantage, not a cost

Treating AI agent infrastructure as an asset your company owns rather than a SaaS service billed per query changes the economics of deploying AI in a small or medium-sized business. Costs become fixed and predictable, data stays under your control, and you can choose an engine such as Hermes Agent, Flowise, or Windmill based on your actual needs rather than what one cloud provider currently offers within its limited subscription plan features.