This page is machine-translated from French. Read the French original.

MCP Servers

THE Model Context Protocol (MCP) is an open standard describing how an AI application discovers and uses external tools. WivenLLM is a client of this standard: any MCP server can be connected, and its tools become skills usable by your agents.

What is it for?

Without MCP, connecting WivenLLM to an internal tool requires specific development. With MCP, the tool simply needs to expose a standard-compliant server: WivenLLM queries the server, discovers the available functions, and offers them to the agents.

The advantage: the MCP ecosystem is shared. A server written for another application works here, and vice versa.

Examples of what is connected like this: a ticketing system, an ERP, a project management tool, an internal service, a proprietary document database.

Configure — administrator

Réglages → Agents, MCP servers section.

An MCP server can be declared in two ways:

Local server. WivenLLM starts the program and communicates with it directly. The command to be executed, its arguments, and its environment variables are declared.

Remote server. WivenLLM connects to an HTTP address. The URL and, if applicable, the authentication headers are declared.

Once declared, the server is started on demand and its tools appear in the skills list, ready to be activated per workspace.

Life cycle

• The servers are started on demand, when an agent needs it.
• The administration screen displays their state and allows them to be restarted.
• A server is unavailable does not prevent the agents to function: the corresponding tools are simply absent.

Points to consider

An MCP server runs with the permissions you give it. A local server runs on your machine; a remote server receives what the agent transmits to it. Only install servers whose origin you know.

The data is being transferred. A remote MCP server receives the arguments sent to it by the agent. In an installation where no data should be transmitted, only local or internal servers are suitable.

Every tool consumes context. A server with fifty different features significantly complicates agent selection. Opt for more focused servers and leverage intelligent skill selection. → SKILLS

The errors are visible. An MCP tool that fails sends its error back to the agent, who reports it — rather than simulating a success.

MCP or agent flow?

The three combine: a flow can call an MCP tool by following a procedure described in a sheet.