fleet

The problem

Ask your agent to train a model and it starts on your laptop — while a 4090 sits idle across the room and a rented A100 bills you by the hour. It cannot use what it cannot see.

• Your agent is stuck on one machine. It has no idea your other boxes exist.
• Finding a free GPU is manual. SSH into five hosts, run nvidia-smi, compare in your head.
• It cannot see what is running, or what is already there — whose job is on that card, which box already serves Llama 3, where the datasets live — so it downloads and installs things twice.

[A 40-second demo. Asked to train on whatever has a free 24 GB card, the agent runs fleet ls --tag cuda --tag vram-24g, picks the idle RTX 4090 over two paid rentals and a busy box, starts train.py there detached, and confirms the card is now in use. Asked what is costing money, it flags an idle H100 rental at $2.49 an hour and leaves stopping it to you.]

Get started: ask your agent

Paste this into Claude Code, Codex, Gemini CLI, or any agent that can run commands, on the machine you work from:

Install fleet from https://github.com/sinolion/fleet

_{Agents: the steps are in INSTALL.md.}

That is the whole setup. fleet is installed once on this machine, and every supported agent you have here learns it — not only the one you asked. From here on you say what you want in plain words, with no commands to remember.

Then add your machines

Tell the agent how you reach a machine — the SSH command you already use is enough — or let it look in your ~/.ssh/config and on your tailnet for machines you can already reach:

You: add my GPU server, ssh [email protected]

Agent: Added as gpu-box: 2× RTX 4090, both idle, 46 GB free. It's ready to use.

Each one takes a sentence, whatever the machine is:

Nothing is installed on the machines you add: they only need SSH, on Linux, macOS or Windows. fleet probes each one for its GPUs, memory and disks, and keeps that current.

Then let the agent pick

You describe the work; the agent finds where it fits. It checks what each machine has and what is free on it right now, so it will not send a job to a busy card:

You: train train.py on whatever has a free 24 GB card

Agent: rtx4090 has 23.1 GB free and an idle GPU; a100-spot is free too but costs $1.89/hr. Starting on rtx4090, logging to train.log.

Keys and passwords never go into the conversation, and anything irreversible waits for your yes.

Only install fleet where you use it. The machines you add need nothing but SSH. Install fleet on another machine only if you also run agents there; it works the same everywhere, with nothing to join or sync. There, fleet ls --discover lists the machines in your ssh config and on your tailnet; the ones you added by hand are in one file you can share with your dotfiles.

An app that cannot run commands?

Desktop apps and editors get fleet in one click; it runs as an MCP server:

Plus plugins for Claude Code, Codex, Copilot CLI and Gemini CLI, and OpenCode, Amp, Windsurf, Cline, Zed, Qwen Code, Goose, Kiro, Hermes — 35+ agents in all: every agent →

Built to be safe

• Nothing to install on your machines. fleet probes with one script over one SSH connection — Linux, macOS or Windows. A NAS or a fresh rental works as it is. The few files it does leave there are listed in What fleet changes.
• No keys handed out, no access managed. fleet connects exactly as you do — your ssh config, your keys and agent, Tailscale SSH, your rental provider's key — and writes no keys anywhere. Who may reach what stays where you already decide it. When a machine refuses, fleet doctor tells you the fix; fleet never makes it.
• No passwords. fleet never asks for one or stores one, and the agent never sees a credential. Secrets in a GPU job's command line are blanked before anyone sees them.
• The agent asks, not guesses. It is told to ask which machine you mean, to leave irreversible commands to you, and not to start work on a GPU someone else is using.
• Nothing phones home. Apart from installing and updating itself, fleet talks to your machines and nothing else: no account, no telemetry, no update check. Found a flaw? SECURITY.md says how to report it privately.

Prefer the command line?

Everything the agent does is a plain fleet command. To install fleet yourself, run this on the machine you work from (this is also what your agent runs when you ask it):

curl -LsSf https://raw.githubusercontent.com/sinolion/fleet/main/install.sh | sh

_{Windows: powershell -ExecutionPolicy ByPass -c "irm https://raw.githubusercontent.com/sinolion/fleet/main/install.ps1 | iex"}

Then drive it yourself:

fleet ls                            # every machine, with what is free right nowfleet ls --tag cuda --tag vram-24g  # by capability: NVIDIA, a card of 24 GB or morefleet ls --tag rocm                 # every AMD machinefleet top                           # live view, like htop for the whole fleetfleet show gpu-box                  # one machine in detail: jobs, software, models, foldersfleet ps                            # every job on every GPU, and which GPUs are idlefleet models                        # every model being served, and how to reach itfleet edit gpu-box --path models=/data/models   # where that machine keeps its modelsfleet ssh gpu-box -- nvidia-smi     # run something therefleet add "ssh [email protected]"     # add a machinefleet ls --discover                 # machines in your ssh config and on your tailnetfleet doctor nas                    # why a machine will not let you in, and the fix

Where a machine keeps its models, datasets and media is written on the machine itself, in machine.yaml (fleet edit NAME --path models=/data/models). Whoever looks at that machine reads the same answer, and the folder sizes and free space come with every probe.

Rentals from vast.ai, RunPod or AutoDL show their price (fleet edit a100 --cost 1.89), the fleet's burn rate, and an alert when a rental sits idle. On Tailscale, ZeroTier or WireGuard? fleet just needs an address it can route to.

How fleet compares

fleet is about the machines you already have. It complements tools that launch new ones.

When to use something else. To rent GPUs on demand at the lowest price, use SkyPilot or dstack: they launch and stop machines, and fleet does not. For weeks of metrics and paging on a large cluster, use DCGM with Prometheus and Grafana. For a live dashboard of many accelerator kinds (Gaudi, NPUs), use all-smi. fleet is for the machines you already have, and for letting your agent use them.

Why not just Tailscale, Ansible, or an ~/.ssh/config? Use them with fleet, not instead of it. Tailscale connects your machines, and fleet runs over it happily, but it does not tell your agent which machine has a free GPU. Ansible changes machines to match a playbook you write, and has no idea what is busy right now. An ~/.ssh/config names your hosts, and fleet reads it. fleet is the part in between: what you have, what is free on it, what is running and installed there, in a form your agent can use without you.

FAQ

Setting up

Safety

Machines

Learn more

• Getting started — your first fleet, in ten minutes
• Guides — adding machines, reaching them, what is on each one, Windows
• Every agent — install commands and config for 35+ agents
• Command reference — every command and option
• How fleet works — what happens, step by step, under the hood
• Design — one core per machine, a skill per agent, MCP for the rest
• All documentation, also as a website: agents-fleet.dev

Status: v0.6. The test suite runs on Linux, macOS and Windows, and every command is run end to end on a real machine of each OS in CI — from a script, as an agent runs it, and at a real terminal, as you do.

Issues and pull requests are welcome — see CONTRIBUTING.md. If fleet saved you a GPU-hour, a ⭐ helps other people find it.

MIT