So MCP does not replace an API. If your CRM has a good API but no MCP server, an agent can use that API directly. Hermes does that when a task calls for it.
What does MCP do for your business?
The biggest benefit is reuse. If an MCP server already exists for your CRM, your project software or your database, nobody has to build the integration from scratch. If you later switch to a different AI application, that server still works.
You also get clearer boundaries. Each MCP server is a separate connection with its own key. You decide per system what the agent may do, and you can switch off one connection without touching the rest.
And an agent can carry out one task across several systems. For a quote request, it looks up the customer in your CRM, reads earlier agreements in your documents and creates a task in your project software. Which integrations make sense and which ones to skip is covered in connecting AI to your CRM and CMS. How we set up these integrations for businesses is described on the AI automation page. You capture how the agent handles such a task, step by step, in a skill. Read more in agent skills.
How does Hermes use MCP?
Hermes, the open-source AI agent Voltti works with, supports MCP out of the box. You add an MCP server to the configuration. At startup, Hermes discovers that server’s tools and registers them. The Hermes documentation on MCP explains how you choose, per server, which tools the agent gets to see. You either list the tools you allow, or list the tools you switch off. For a payment system, that means you can allow looking up payments but not issuing refunds. A local MCP server also does not get all your environment variables by default, only what you configure plus a safe baseline.
Our own Hermes installation at Voltti works this way with Storyblok, the CMS behind this website, and with Search Console. Each system has its own access. Search Console is read-only: the agent pulls search data from it but cannot change anything there. In Storyblok, the agent edits content and in some cases publishes it too. We review the work afterward and step in where needed. To see what that looks like in practice, read retrieving marketing data with an AI agent and SEO with an AI agent.
What are the risks of MCP?
MCP is a protocol. How safe an integration is depends on the server and on the permissions you grant. Pay attention to these four points:
Permissions that are too broad. A server with a key that can do everything gives the agent everything. Create a separate key for each server with only the permissions the task needs.
Hidden instructions. A document, email or web page the agent reads through a server can contain instructions. That is called prompt injection. Read more in prompt injection and API keys.
Third-party code. An MCP server is software that runs on your systems. Only install servers whose source you know, and check what a server runs during installation.
Too many tools. Every tool the agent sees is a possible action. Switch off tools you do not need, and have a person approve actions with major consequences first.
The official documentation has a separate page with security best practices for MCP, mainly for people who build servers. For a practical list for your own business, see our safety checklist for AI and business data. How we protect data and access is explained on the security and privacy page.
What if your software has no MCP server?
Then there are other routes. If the software has an API, the agent can use it. If there is no API and no MCP server, we build a custom integration ourselves. The form depends on what the system allows. It can be a small connector with its own API endpoint. The agent can also read exports, such as a CSV file or a scheduled report. Or the agent operates the web interface itself, through browser automation. Even then, the agent only gets the permissions the task needs.
Where do you start?
Start with the work. Pick one task that comes back every week and look at which systems the agent needs for it. Only then do you choose the route: an existing MCP server, the API or a custom integration. In a pilot, you test that one integration alongside your current way of working. A pilot usually takes 2 to 4 weeks, from the first conversation to the evaluation. How that works is described in our approach.