Fleet

What fleet does

The fleet MCP server connects a coding agent to a set of machines reachable through SSH. It keeps the machine list at a central host and reports current resource information such as GPUs, free VRAM, CPU, RAM, disk space, operating system, and running GPU processes. This lets an agent select a host based on available capacity rather than relying only on machine names.

Added hosts can include workstations, GPU servers, laptops, NAS devices, and rented machines. Rental entries can include an hourly cost, allowing fleet to show burn rate and identify paid machines that are idle. The machines being managed do not need fleet installed; they need SSH access and a supported operating system such as Linux, macOS, or Windows.

How it works

Install fleet on the machine where you use your coding agents. That machine acts as the center for one fleet. Other machines are added with the SSH command or host configuration you already use. fleet probes each host through SSH and refreshes its hardware and availability data.

A second machine can join as a fleet member through an invite line if agents also run there. The center controls which machines may connect to one another, and access changes can be allowed or revoked with fleet commands. Credentials are not passed through the agent conversation. If a password is needed, the user enters it once, and fleet does not store it.

The fleet MCP server can be used through supported agent plugins and MCP application integrations. It does not require an account or cloud service, and the project states that it does not send telemetry or perform update checks.

Setup and configuration

Install on Linux or macOS with the project’s shell installer, or use the documented PowerShell installer on Windows. An agent can also be instructed to install from the repository. After installation, add machines with an SSH command, including optional identity files, ports, or configured SSH host names.

Use an invite when the center cannot directly reach a machine, or generate a key for a cloud-init template when provisioning a machine that does not yet exist. Rental pricing can be recorded with the fleet edit command. The project documents the files and SSH changes made on managed machines, including how to remove them.

Tools and capabilities

The fleet MCP server supports capabilities represented by commands including:

• fleet ls for listing machines and current availability.
• fleet ls --tag ... for filtering by capabilities such as CUDA or VRAM size.
• fleet top for a live, fleet-wide resource view.
• fleet show NAME for host details, GPU processes, services, and disks.
• fleet ssh NAME -- COMMAND for running a command remotely.
• fleet add and fleet invite for adding machines or creating join instructions.
• fleet access for allowing or denying machine-to-machine access.
• Rental cost tracking and idle paid-machine alerts.

Irreversible actions are intended to wait for user confirmation, while routine discovery and status checks can be handled by the agent.

Limitations and notes

The project requires SSH connectivity or an invitation-based join path. It is not a hosted inventory service, and the provided material does not describe support for machines that cannot be reached through either method. fleet must be installed on each center or member machine where agents need to use it; ordinary managed hosts only need SSH.

The README names integrations for Claude Desktop, Cursor, VS Code, Claude Code, Codex, Gemini CLI, and several other agent applications. Availability and installation steps vary by client, so use the client-specific guidance in the project documentation.