Local MCP

The Local MCP tool allows a Model Context Protocol (MCP) server to be run locally, with its tools selectively made available to agents.

Prerequisites

Configure a code environment with the following specifications:

  • Python version 3.10 or higher

  • The fastmcp package version 2.0 or higher

  • Containerized execution and AlmaLinux 9 or higher are recommended. The Agent MCP server runtime addition installs uvx and npx to help start most MCP servers.

Configuration

Create a tool with type “Local MCP”. Then the configuration of the tool involves two main steps:

  • Define the command, arguments, and environment variables used by the MCP server process in the tool’s settings.

Note

Most MCP servers provide a JSON configuration file/example. Fill the configuration automatically using the Paste config button.

  • The Load tools button loads the list of tools exposed by the MCP server. Each tool must then be enabled individually to be made available to agents.

Usage

Add the Local MCP tool to an agent like any other tool. When running, the agent directly sees and uses each enabled MCP server tool as a regular standalone tool.

Security

The Local MCP tool runs under the querying user’s identity (see User Isolation) and therefore with that user’s permissions / accesses.

An administrator may disable the creation/edition of Local MCP Server tools, or restrict it to only administrators in Administration > Settings > LLM Mesh.

Some MCP server runners, such as uvx and npx, download the server’s code every time they start. To control updates, you can pin the version of the MCP server.