AgentStack
Browse Sign in
Browse Why AgentStack Sell Docs
Sign in
MCP unreviewed Apache-2.0 Self-run

Agent Fleet

mcp-metahub-tech-agent-fleet · by metahub-tech

CLI wizard + MCP servers giving an LLM agent its own fleet of physical devices (Windows / macOS / Linux / Android, iOS planned). One command: uvx agent-fleet setup.

No reviews yet
0 installs
34 views
0.0% view→install

Install

$ agentstack add mcp-metahub-tech-agent-fleet

Open-source listing, not yet scanned by AgentStack. Follow the source repository for install instructions.

Reliability & compatibility

Not yet reviewed
0 installs to date
no reviews yet
2mo ago

Declared compatibility

Claude CodeClaude DesktopCursorWindsurf

Compatibility is declared by the source manifest. End-to-end runtime verification is coming, see below.

Preview Execution monitoring

We're building live execution health for every listing: tool-call success rate, median latency, uptime, and last-checked timestamps, measured, not self-reported. It isn't live yet, so we don't show numbers we can't stand behind.

How agent discovery & health will work →
Are you the author of Agent Fleet? Claim this listing to set pricing, connect Stripe payouts, and keep 70% of every sale.
Sign up to claim

About

agent-fleet

[](LICENSE) [](https://github.com/metahub-tech/agent-fleet/releases/tag/v0.8.4-alpha)

English · [简体中文](README.zh-CN.md) · [日本語](README.ja.md) · [한국어](README.ko.md) · [Español](README.es.md) · [Français](README.fr.md) · [Deutsch](README.de.md) · [Português (BR)](README.pt-BR.md) · [Русский](README.ru.md)

An LLM agent driving a real iPad over MCP — no human hands.

> Give your LLM agent its own fleet of physical devices. > Connect real Windows / macOS / Android / iOS hardware to your LLM agent over MCP so it can operate them like a human. One command to install: > > ``bash > uvx --from "git+https://github.com/metahub-tech/agent-fleet@v0.8.4-alpha#subdirectory=cli" agent-fleet setup > ``

What is it

Bring real devices — Windows PCs, Macs, Android phones, iPhones — into your LLM agent's toolchain so it can drive them just like local commands: take screenshots, tap buttons, run tests, read logs, debug GUIs.

This is infrastructure for agent-driven software testing and cross-platform verification — and, more broadly, for giving agents real perception of and leverage over the physical world.

┌─────────────────┐                              ┌──────────────┐
│  Agent (any OS) │ ──── Tailscale Mesh ─────>  │ Windows PC   │ win-device     :8766 ✅
│  Claude Code /  │                              ├──────────────┤
│  Cursor / Cline │                              │ macOS box    │ mac-device     :8767 ✅
│  / OpenClaw /   │                              ├──────────────┤
│  Antigravity /  │                              │ Android phone│ android-device :8768 ✅
│  Hermes / ...   │                              ├──────────────┤
└─────────────────┘                              │ iPhone/iPad  │ ios-device     :8769 ✅
                                                 └──────────────┘
            ↑                                                ↑
   uvx agent-fleet setup           generates configs for 6 agent frameworks
   one command: install client + server / configure Tailscale / guide permissions / self-check / emit snippets

Status

| Component | Version | Status | |---|---|---| | Windows 10/11 bridge | 0.2.0 | ✅ Released (win-device, 74 tools — 43 core+discovery, +27 agentbrowser, +1 humanbrowser, +3 vision (browser caps need Chrome/node); streamable-http) | | macOS 12+ bridge | 0.3.0 | ✅ Released (mac-device, launchd, 73 tools — 42 core+discovery, +27 agentbrowser, +1 humanbrowser, +3 vision (browser caps need Chrome/node); GUI-permission flow) | | Android bridge | 0.7.0-alpha | ✅ Released (android-device, 31 tools, multi-device + USB + Wireless + Hybrid ADB) | | agent-fleet CLI wizard | 0.5.0-alpha | ✅ Released (uvx agent-fleet setup one-shot install; configs for 6 frameworks) | | role rename → -device + macOS permission primer | 0.6.0-alpha | ✅ Released | | v0.6.x patches (UI introspection, smoke tests, bugfixes, installer hardening…) | 0.6.1–0.6.15 | ✅ Released — see [CHANGELOG.md](CHANGELOG.md) | | iOS / iPadOS bridge | 0.8.0-alpha | ✅ Released (ios-device, 30 tools, WebDriverAgent + pymobiledevice3, iPad verified) | | iOS WDA daemon (boot-survival + keep-alive) | 0.8.2-alpha | ✅ Released (go-ios runwda + tunneld launchd; free/paid signing modes) | | Cross-device coordination | 0.10.0 | 🔭 Future | | Public stable release | 1.0.0 | 🔭 Future (after community feedback on the alpha) |

See [docs/roadmap.md](docs/roadmap.md).

Quick Start

Newcomers / one-shot flow — on the device to be controlled (a PC, or a PC with phones attached):

# macOS / Linux  —  NOT `curl ... | bash`!  The wizard is interactive and bash
# piped from curl uses stdin for the script itself, so questionary's prompts
# die with EOFError.  The `bash -c "$(...)"` form puts the script in argv and
# leaves stdin = terminal.
bash -c "$(curl -fsSL https://raw.githubusercontent.com/metahub-tech/agent-fleet/main/install.sh)"

# Windows  —  `irm | iex` is fine on PowerShell because iex runs the script in
# the current session and Read-Host reads from the console host, not stdin.
powershell -ExecutionPolicy Bypass -c "irm https://raw.githubusercontent.com/metahub-tech/agent-fleet/main/install.ps1 | iex"

Or, if you already have uv (during the v0.8.4-alpha phase it is pulled from git, not yet on PyPI):

uvx --from "git+https://github.com/metahub-tech/agent-fleet@v0.8.4-alpha#subdirectory=cli" agent-fleet setup

> macOS 12 users must run brew install coreutils first (uv's wrapper uses realpath, which macOS 12 lacks by default; see #2).

The wizard walks you through: pick a role → install the MCP server → configure Tailscale → interactively guide GUI permissions / ADB authorization → run a health check → emit config snippets for 6 agent frameworks.

Power users can still call the underlying scripts directly — docs/install-pattern.md remains valid (the "advanced manual").

| You are | Read | |---|---| | A newcomer (let the wizard drive) | run the command above and follow the prompts | | A device admin scripting your own deployment | [docs/install-pattern.md](docs/install-pattern.md) (low-level script contract) | | An agent operator | paste the wizard's snippet into your agent config; see also [docs/agent-host-setup.md](docs/agent-host-setup.md) | | A contributor (design docs) | [docs/internal/design/2026-05-11-agent-fleet-cli.md](docs/internal/design/2026-05-11-agent-fleet-cli.md) |

Architecture

flowchart LR
  A["🤖 LLM agentClaude Code · Cursor · Cline · …unified MCP tools"]
  A ==>|"① connects to its own computerMCP over Tailscale (WireGuard, cross-LAN)"| HUB
  subgraph HUB["🖥️ The computer = the agent's hands · Windows / macOS"]
    direction TB
    SELF["drive the computer itselfwin-device :8766 · mac-device :8767pywinauto · AppleScript · shell"]
    JUMP["② a jump host to attached devices"]
  end
  JUMP -->|"USB · adb"| AND["📱 Androidandroid-device :8768"]
  JUMP -->|"USB · WebDriverAgent"| IOS["📱 iPhone / iPadios-device :8769"]
  JUMP -.->|"planned · hdc"| HM["📱 HarmonyOS(planned)"]
  JUMP -.->|"③ anything a human can manage via the computer"| MORE["⋯ more devices"]
  classDef agent fill:#1f6feb,stroke:#0b3d91,color:#fff
  classDef hub fill:#0e7490,stroke:#053b46,color:#fff
  classDef dev fill:#16a34a,stroke:#064e23,color:#fff
  classDef plan fill:#6b7280,stroke:#374151,color:#fff
  class A agent
  class SELF,JUMP hub
  class AND,IOS dev
  class HM,MORE plan

The agent connects to a computer (Windows/macOS) — its "hands" — and uses it both to drive the computer directly and as a jump host to the devices attached to it: Android and iOS today, HarmonyOS next. In principle, any device a human can manage through a computer, the agent can too. (The iOS host must be a Mac.)

The tool interface stays semantically consistent across platforms (take_screenshot / tap / launch_app / …); switching devices is just swapping a URL in ~/.claude.json.

See [docs/architecture.md](docs/architecture.md).

Repository layout

agent-fleet/
├── docs/                          # shared docs
│   ├── install-pattern.md         # developer baseline: two roles / two install paths / directory contract
│   ├── architecture.md            # common bridge architecture
│   ├── agent-host-setup.md        # agent-side config (~/.claude.json + skill symlinks)
│   ├── roadmap.md                 # platform roadmap
│   └── platforms/.md        # per-platform manual (device side)
├── platforms/                     # one self-contained bay per platform
│   ├── windows/
│   │   ├── platform.toml          # manifest: id / port / host_os / enabled capabilities
│   │   ├── README.md              # at-a-glance
│   │   ├── server/                # MCP server source + deps
│   │   ├── scripts/               # install / launch / troubleshoot scripts
│   │   ├── skills/using-win/      # skill docs for the agent to use
│   │   └── examples/              # claude-settings.json snippets
│   ├── macos/                     # same sub-structure
│   ├── android/                   # same sub-structure
│   └── ios/                       # same sub-structure
├── scripts/                       # repo-level scripts
│   └── install-agent-side.py      # one command to install MCP + skill into ~/.claude.json
├── examples/                      # cross-platform examples
│   └── multi-platform-claude-settings.json
└── CHANGELOG.md

To add a platform → drop the same sub-structure under platforms//. See [docs/install-pattern.md § Adding a new platform](docs/install-pattern.md).

License

[Apache 2.0](LICENSE)

Contributing

Contributions welcome! agent-fleet is publicly released under the Apache 2.0 license and accepts community contributions.

See [CONTRIBUTING.md](CONTRIBUTING.md) for the contribution flow and coding conventions, and [CODE_OF_CONDUCT.md](CODEOFCONDUCT.md) for the code of conduct. Report security vulnerabilities privately via the GitHub Security tab → "Report a vulnerability".

Source & license

This open-source MCP server is cataloged on AgentStack and links to its original source — we do not rehost the code.

Install and usage instructions live in the source repository linked above.

Reviews

No reviews yet, be the first.

Versions

  • v0.1.0 Imported from the upstream source.