Linux MCP Server

The Linux MCP Server connects AI assistants such as Claude Code, Cursor, or Claude Desktop to Linux systems so that they can narrow down faults there. Instead of typing commands one by one and pasting the output into a chat by hand, you ask the assistant something like "Why is the web server on this host slow?" or "Which services have failed?". The assistant then calls tools of the server, reads system information, service status, processes, logs, network, and storage, and summarizes the findings. The project belongs to rhel-lightspeed from Red Hat and, according to the repository, is licensed under Apache-2.0.

What is the Linux MCP Server?

MCP, the Model Context Protocol, is an open protocol through which AI assistants call tools of external services. The Linux MCP Server is such a service for system administration. According to the project page and the package description it is designed for read-only diagnostics and troubleshooting and intended for Red Hat Enterprise Linux, Fedora, CentOS Stream, and other systemd-based distributions. Red Hat lists it in the RHEL 9.8 release notes as a Developer Preview, that is a preview without the full scope of support of a regular product feature. Anyone using it in production should be aware of that.

Which tools are there?

The documentation groups the tools into areas: system information, services, processes, logs, network, storage, Performance Co-Pilot, and script execution. With them you can, for example, inspect system state and hardware, check services and processes, read journal and log files, look at network connections, and evaluate storage usage. Performance data is reachable through Performance Co-Pilot. Several hosts can be addressed in one session, so a problem can be compared across multiple servers.

What is it particularly useful for?

The greatest benefit lies in incident analysis. Anyone who looks after Linux servers knows the pattern: a service does not start, the cause is somewhere in the journal, and getting there takes several commands. The assistant takes over this sequence, knows the common failure patterns, and explains what it sees. Compared with a general shell connection, the access is more narrowly scoped because, according to the documentation, the default tools only read. That makes use on production systems easier to oversee, but it does not replace checking the permissions of the account in use.

Setup

A prerequisite is a Python tool manager: according to the project page you first install uv and then the server with uv tool install linux-mcp-server. After that you enter it in your client's MCP configuration, in ~/.claude.json for Claude Code and in ~/.cursor/mcp.json for Cursor. The command is the path to the installed program, and an environment variable such as LINUX_MCP_USER sets the SSH user. According to the documentation there is alternatively a container image for Podman or Docker. Connections to remote hosts run over SSH with key authentication; passwords are not intended.

Security and limits

The server acts with the permissions of the SSH account on the target system, so use an account with as few permissions as possible. Access to log files is limited by the environment variable LINUX_MCP_ALLOWED_LOG_PATHS to a list of permitted paths; without this variable the tool cannot read any files. Host key verification is on by default and should stay on. The documentation warns that the HTTP transport offers no authentication and should not run in production or on untrusted networks; the default is stdio. The toolset for script execution differs from the read-only default and is only enabled with the start option --toolset run-script or --toolset both. According to the documentation it can change system state and uses validation as well as confirmation. Leave it off if you only want to diagnose. What logs and command output contain enters the AI client context and therefore reaches the model provider; the documentation describes no separate data retention for this.

Costs

The project is open source and free to use. Whether a subscription is needed for use with RHEL and which scope of support applies is not stated on the pages that could be retrieved; Red Hat's documentation is authoritative. Costs of your AI client and the model provider come in addition.

Who is the Linux MCP Server for?

It is worthwhile for administrators and development teams who run Linux servers and spend time searching logs during incidents. Anyone who runs only Windows or container platforms without SSH access gains little from it. Because of its preview status it is initially suited to test and development systems and to read-only diagnostics before it is used in critical environments.

The Python tool manager uv (or a container runtime), an MCP client that supports local stdio servers, and for remote hosts SSH access with key authentication.

With the Linux MCP Server, an AI assistant diagnoses Linux systems read-only: services, processes, logs, network, and storage. This guide leads from installation to the first query.

Step 1: Check the prerequisites

You need the Python tool manager uv, an MCP client that supports local servers, and for remote hosts SSH access with a key. Note that Red Hat lists the server as a Developer Preview; use it on test and development systems first.

Step 2: Install

Install uv following the instructions on docs.astral.sh and then the server with uv tool install linux-mcp-server. The command usually places the program at ~/.local/bin/linux-mcp-server; you need this path in a moment.

uv tool install linux-mcp-server

Prepare SSH

Connections to remote hosts run over SSH with a key; passwords are not intended. Set up a key and define in ~/.ssh/config which hosts are reachable. Keep host key verification on. Use an account with as few permissions as possible, because the server acts with its permissions.

Step 3: Add it to Claude Code

In ~/.claude.json, add an entry linux-mcp-server under mcpServers: command is the path to the program, args stays empty, and in env you set LINUX_MCP_USER to your SSH user name. For Cursor the same structure applies in ~/.cursor/mcp.json, for Claude Desktop in the file claude_desktop_config.json; restart Claude Desktop completely afterwards.

Limit log access

Set the variable LINUX_MCP_ALLOWED_LOG_PATHS in env, for example to /var/log/messages,/var/log/lastlog. According to the documentation the log tool cannot read any files without this variable. List only paths you actually need.

Step 4: First queries

Start read-only: "Show system information for this host.", "Which services have failed?", or "Which processes use the most memory?". Name the host as precisely as possible if you have configured several.

Decide on script execution deliberately

The script execution tools are off by default and only enabled with the start option --toolset run-script or --toolset both. According to the documentation they can change system state. Leave them off as long as you only diagnose, and enable them only on systems where you review every confirmation. Do not use the HTTP transport in production; it offers no authentication.

Passwordless SSH key authentication to the target hosts; in default operation (stdio) the server itself has no sign-in of its own. According to the documentation the HTTP transport offers no authentication.

Read access to system information, services, processes, logs (permitted paths only), network, storage, and performance data within the permissions of the SSH account; optional and only enabled on request: script execution, which can change the system.

The server runs locally on the user's machine and connects to the target systems over SSH. Output of logs and commands enters the AI client context and therefore reaches the model provider. The documentation describes no data retention of its own.

Overly broad permissions of the SSH account, exposure of sensitive logs to the model provider, unprotected HTTP transport, disabled host key verification, and, if enabled, script execution that can change system state.