Agents¶

Most people never need this page. Asking an agent to install fleet — or running the installer — sets up every supported agent on the machine at once, with fleet setup. This page is for adding fleet from inside one particular agent or app, or to an agent fleet setup does not know.

Every way in runs the same fleet from PyPI (agents-fleet), so all your agents on a machine share one fleet: one list of machines, one cache (why). Two kinds of entry:

Skills are for agents that can run commands. A skill is text that tells the agent how to use the fleet command, and costs nothing until a task needs a machine. It needs the fleet command installed; if it is missing, the skill says how to get it.
MCP is for apps that cannot run commands. The app launches uvx agents-fleet mcp, which runs the same commands for it. It needs only uv.

However it is installed, it is the same fleet. There are no modes to choose. The first use of fleet on a machine — through a plugin, an MCP app or a shell — records the machine and teaches the other agents found there, once (FLEET_NO_FIRST_RUN=1 skips that). fleet then reaches other machines with your own ssh, exactly as you do.

✓ = run here, and the agent registered (or connected to) the server. docs = from the agent's official documentation, not run here.

Plugins and extensions — the skill, installed by the agent
One command — mcp add and similar
One click — install links and bundles
Config files — everything else
Skills only and what fleet setup does

Plugins and extensions¶

The agent installs fleet's skill, which teaches it the fleet commands and costs nothing until a task needs a machine.

Agent	Install
Claude Code	`/plugin marketplace add lion-zhang/fleet` then `/plugin install fleet@fleet` (shell: `claude plugin marketplace add lion-zhang/fleet && claude plugin install fleet@fleet`)	✓
Codex	`codex plugin marketplace add lion-zhang/fleet && codex plugin add fleet@fleet`	✓
GitHub Copilot CLI	`copilot plugin marketplace add lion-zhang/fleet && copilot plugin install fleet@fleet`	✓
Gemini CLI	`gemini extensions install https://github.com/lion-zhang/fleet` — the skill and the MCP server	✓

One command¶

These register fleet's MCP server, launched as uvx agents-fleet mcp.

Agent	Command
Claude Code	`claude mcp add --scope user fleet -- uvx agents-fleet mcp`	✓
Codex (CLI, IDE, ChatGPT app)	`codex mcp add fleet -- uvx agents-fleet mcp`	✓
GitHub Copilot CLI	`copilot mcp add fleet -- uvx agents-fleet mcp`	✓
Amp	`amp mcp add fleet -- uvx agents-fleet mcp`	✓
OpenCode	`opencode mcp add fleet -- uvx agents-fleet mcp`	✓
Qwen Code	`qwen mcp add fleet uvx agents-fleet mcp`	✓
Auggie (Augment CLI)	`auggie mcp add fleet -- uvx agents-fleet mcp`	✓
Factory Droid	`droid mcp add fleet "uvx agents-fleet mcp"`	✓
Kiro CLI	`kiro-cli mcp add --name fleet --scope global --command uvx --args agents-fleet --args mcp`	✓
Mistral Vibe	`vibe mcp add fleet --transport stdio --command uvx --arg agents-fleet --arg mcp`	✓
VS Code (Copilot agent mode)	`code --add-mcp '{"name":"fleet","command":"uvx","args":["agents-fleet","mcp"]}'`	docs
Windsurf / Devin Desktop	`devin mcp add fleet -s user -- uvx agents-fleet mcp`	docs
Jan	`jan cli mcp add fleet --command uvx --arg agents-fleet --arg mcp --active`	docs
Crush	`echo 'mcp add fleet --command uvx --args agents-fleet --args mcp' >> ~/.config/crush/crushrc`	docs

One click¶

Agent	Link
Cursor	Add to Cursor
VS Code	Install in VS Code · VS Code Insiders
Visual Studio	Install in Visual Studio
Kiro	Add to Kiro
LM Studio	Add to LM Studio
Goose	`goose://extension?cmd=uvx&arg=agents-fleet&arg=mcp&id=fleet&name=fleet&description=fleet&timeout=300` (paste into a browser)
Claude Desktop	download `fleet.mcpb` from the latest release and open it ✓

These links are built from each agent's documented format; the Claude Desktop bundle was validated, packed and run here.

Config files¶

For everything else, add fleet to the agent's config. Most agents read the same shape:

{
  "mcpServers": {
    "fleet": { "command": "uvx", "args": ["agents-fleet", "mcp"] }
  }
}

Agent	File (user scope)	Notes
Cursor (IDE and CLI)	`~/.cursor/mcp.json`
Windsurf / Devin Desktop	`~/.config/devin/mcp_config.json` (Windows `%APPDATA%\devin\`); the older `~/.codeium/windsurf/mcp_config.json` is still read
Claude Desktop	macOS `~/Library/Application Support/Claude/claude_desktop_config.json`, Windows `%APPDATA%\Claude\claude_desktop_config.json`
Cline (VS Code)	MCP Servers → Configure → Configure MCP Servers (`cline_mcp_settings.json`)	add `"disabled": false`
Cline CLI	`~/.cline/data/settings/cline_mcp_settings.json`	add `"disabled": false`
Kiro (IDE and CLI)	`~/.kiro/settings/mcp.json`
GitHub Copilot CLI	`~/.copilot/mcp-config.json`
Qwen Code	`~/.qwen/settings.json`
Factory Droid	`~/.factory/mcp.json`
Auggie	`~/.augment/settings.json`
Junie (IDE and CLI)	`~/.junie/mcp/mcp.json`
JetBrains AI Assistant	Settings → Tools → AI Assistant → Model Context Protocol → + → As JSON	paste the block above
Augment Code (IDE)	Augment settings → MCP → Import from JSON	paste the block above
Kimi Code	`~/.kimi-code/mcp.json`
Warp	`~/.warp/.mcp.json`
LM Studio	Program → Install → Edit mcp.json
Trae	Settings → MCP → Add → Add Manually	paste the block above
Cherry Studio	Settings → MCP Servers → Add → import from JSON
BoltAI	`~/.boltai/mcp.json`
Jan	Settings → MCP Servers → +	add `"active": true`
Msty Studio	Toolbox → Tools → Add New Tool → STDIO / JSON	the inner object only: `{"command": "uvx", "args": ["agents-fleet", "mcp"]}`

The agents below use their own shapes.

VS Code — user mcp.json (MCP: Open User Configuration), or .vscode/mcp.json in a project

{ "servers": { "fleet": { "type": "stdio", "command": "uvx", "args": ["agents-fleet", "mcp"] } } }

Visual Studio reads the same shape from `%USERPROFILE%\.mcp.json`.

OpenCode (~/.config/opencode/opencode.json) and Kilo Code (~/.config/kilo/kilo.json) — ✓

{ "mcp": { "fleet": { "type": "local", "command": ["uvx", "agents-fleet", "mcp"], "enabled": true } } }

Zed — ~/.config/zed/settings.json (Windows %APPDATA%\Zed\settings.json)

{ "context_servers": { "fleet": { "command": "uvx", "args": ["agents-fleet", "mcp"], "env": {} } } }

Continue — ~/.continue/config.yaml (agent mode only)

mcpServers:
  - name: fleet
    type: stdio
    command: uvx
    args: ["agents-fleet", "mcp"]

Goose — ~/.config/goose/config.yaml (Windows %APPDATA%\Block\goose\config\config.yaml)

extensions:
  fleet:
    name: fleet
    type: stdio
    cmd: uvx
    args: [agents-fleet, mcp]
    enabled: true
    timeout: 300

Codex — ~/.codex/config.toml

[mcp_servers.fleet]
command = "uvx"
args = ["agents-fleet", "mcp"]

Amp — ~/.config/amp/settings.json

{ "amp.mcpServers": { "fleet": { "command": "uvx", "args": ["agents-fleet", "mcp"] } } }

Mistral Vibe — ~/.vibe/config.toml

[[mcp_servers]]
name = "fleet"
transport = "stdio"
command = "uvx"
args = ["agents-fleet", "mcp"]

Not supported: Aider has no MCP support, and ChatGPT's chat reaches only remote MCP servers; both can still run fleet as a shell command. Roo Code has shut down; its successor is Kilo Code, above.

Skills only¶

Agents that read SKILL.md files can get fleet's skill without any MCP server:

npx skills add lion-zhang/fleet          # ✓ — installs into every agent it recognises

For Hermes, install the CLI and run fleet setup --target hermes, which writes into Hermes' own home (%LOCALAPPDATA%\hermes on Windows, or $HERMES_HOME).

What the MCP server offers¶

Each tool runs one fleet command and returns what it says, so an app without a shell gets the same answers as an agent with one.

Tool	Runs
`list_machines`	`fleet ls --json`, with `tag` to pick by what a machine has, `fields` for only some keys, `refresh`/`rescan` to measure now
`show_machine`	`fleet show NAME --json`, with `rescan`
`list_gpu_jobs`	`fleet ps --json`, with `refresh`
`list_models`	`fleet models --json`, with `refresh`/`rescan`
`run_on_machine`	`fleet ssh NAME -- COMMAND`: its output and exit code, within two minutes
`add_machine`	`fleet add "ssh …" --json`
`discover_machines`	`fleet ls --discover --json`
`doctor_machine`	`fleet doctor [NAME] --json`
`edit_machine`	`fleet edit NAME`: tags, cost, note (on the machine; `local` keeps tags and cost here), address, disks, name, alias
`set_machine_paths`	`fleet edit NAME --path KIND=PATH`
`set_machine_endpoint`	`fleet edit NAME --endpoint NAME=PORT[/PATH]`: a model server the probe would not find

Removing a machine, and installing, updating or uninstalling fleet, are not tools: they are yours to run, in a terminal.

What `fleet setup` does¶

With the CLI installed (the one-line installer runs this for you), fleet setup teaches every agent it finds on the machine, and touches nothing else — it never creates a config directory for an agent you do not use, and in files you own it edits only its own marked region.

Agent	What it gets
Claude Code	skill: `~/.claude/skills/fleet/SKILL.md`
Codex, Gemini CLI, Copilot CLI, OpenCode, Kilo, Amp	one shared skill: `~/.agents/skills/fleet/SKILL.md` — every one of them reads it (checked with each CLI), so it is written once
Hermes	skill, in Hermes' own home
Claude Desktop, Cursor, VS Code, Windsurf	the MCP server, merged into the app's own config

Older versions put Codex's skill in ~/.codex/skills and a region in ~/.gemini/GEMINI.md; the next fleet update (or fleet setup --refresh) moves them, so no agent ever reads fleet twice.

fleet setup --project writes into the current repository instead: .claude/skills/fleet/SKILL.md, or with --target codex (or hermes, gemini) a marked region in the AGENTS.md / GEMINI.md those agents read. Updating fleet (fleet update) refreshes what every agent is told, on every machine.

Using the installed CLI instead of uvx: put the full path that which fleet prints as the command, with ["mcp"] as the args. Desktop apps on macOS do not see ~/.local/bin on their PATH, so the bare name fleet is not enough there.