Manos

What it is

Manos drives Android emulators/devices (via adb) and iOS simulators (via xcrun simctl, with idb for native UI interaction) over the Model Context Protocol. It's purpose-built for the exploratory, test-free loop where an LLM agent pokes at an app and reacts to what it sees: atomic act+observe actions, control over the device conditions real users hit, crash/log and network capture for when something breaks, an accessibility audit, and one-call recording that promotes an ad-hoc session into a replayable regression test. See the design notes & roadmap for the rationale.

Installation

Manos is a stdio MCP server published to npm. Most AI clients can launch it on demand with npx — no install step — so the snippets below use that. Prefer a resident binary? npm install -g manos-mcp and use a bare manos serve instead.

1 · Install (optional)

# run on demand — nothing to install:
npx -y manos-mcp serve

# …or install the CLI globally:
npm install -g manos-mcp

Requires Node 20+. Working on manos itself, or pinning a local build? Clone the repo, npm install (it builds via the prepare script), and point your client at node /ABS/PATH/to/Manos-MCP/dist/cli.js serve.

2 · Register with your AI client

claude mcp add manos -- npx -y manos-mcp serve

# globally installed binary instead:
claude mcp add manos -- manos serve

{
  "mcpServers": {
    "manos": {
      "command": "npx",
      "args": ["-y", "manos-mcp", "serve"]
    }
  }
}

{
  "mcpServers": {
    "manos": {
      "command": "npx",
      "args": ["-y", "manos-mcp", "serve"]
    }
  }
}

{
  "servers": {
    "manos": {
      "type": "stdio",
      "command": "npx",
      "args": ["-y", "manos-mcp", "serve"]
    }
  }
}

{
  "mcpServers": {
    "manos": {
      "command": "npx",
      "args": ["-y", "manos-mcp", "serve"]
    }
  }
}

command: npx
args:    ["-y", "manos-mcp", "serve"]
transport: stdio

3 · Verify the toolchain

npx -y manos-mcp doctor        # or: manos doctor

doctor reports which backends it found (adb, xcrun, idb, maestro, OCR engine), lists connected devices, and prints each device's live capabilities — so you know what works before you rely on it. Android needs adb (auto-detected from $ANDROID_HOME / the default SDK path); iOS needs Xcode's xcrun. Everything else is optional and only gates the advanced features below.

Basic operations

The everyday unit of work is act → observe: most sessions cycle through a handful of tools, and each action can return the resulting screen so you rarely call inspect_screen twice in a row.

A typical loop

Sign in, then wait for the landing screen — five calls, each act returning what changed:

list_devices                                              # → pick a device_id
inspect_screen   { device_id }                            # compact tree, stable ids
tap              { device_id, text: "Sign In", observe: "diff" }
input_text       { device_id, text: "user@example.com" }
tap              { device_id, text: "Continue", observe: "screen" }
wait_for         { device_id, text: "Welcome", timeout_ms: 8000 }

Android vs iOS capability matrix

Capabilities are reported live per device by the device_capabilities tool and the doctor command, including the active backend — the agent never has to guess what a platform supports.

Advanced operations & setup

These go beyond tap/inspect. Each is capability-gated and degrades gracefully — Manos tells you what's missing instead of failing opaquely. Where a feature needs something special on the developer's machine or build, the setup is called out below.

Tool reference

Upcoming features

The roadmap, drawn from IMPROVEMENTS.md. Status reflects priority, not a commitment date.