# callstackincubator/agent-device [Health: Active]

**Category:** 🖥️ OS Automation  
**Repository:** https://github.com/callstackincubator/agent-device  
**GitHub Stars:** 5164  
**Views:** 6  
**Installs:** 0  
**Upvotes:** 0  
**Directory Page:** https://allmcps.com/mcp/callstackincubator-agent-device

## Description
Discovery router for the agent-device CLI. Exposes status, install guidance, version-matched help, prompts, and resources for iOS, Android, tvOS, macOS, and Linux automation workflows.

## Claude Desktop Quick Installation
No confirmed install details yet, so no config is published. Follow the setup steps in the repository above.

## Documentation

## What callstackincubator/agent-device does

callstackincubator/agent-device connects coding agents to the agent-device CLI for app automation and verification. Its MCP entry point exposes the installed CLI commands as structured tools, allowing an agent to work with running apps instead of relying only on source code or screenshots.

The server supports workflows across iOS, Android, and HarmonyOS simulators, emulators, and physical devices, along with tvOS, Android TV, Fire TV, Amazon Vega OS TV, web, macOS, and Linux. It can provide setup status, installation guidance, version-matched help, prompts, and resources so an agent can discover the commands and workflow guidance appropriate to its environment.

## How it works

Start the official stdio server with the `agent-device mcp` command. MCP-compatible clients can then invoke the same execution path used by the CLI. A typical configuration points the client at the installed `agent-device` executable and passes `mcp` as its argument.

The underlying workflow is session-based. An agent opens an application, captures an accessibility snapshot, and uses references or selectors to interact with visible elements. Actions can include pressing controls, filling fields, scrolling, gesturing, waiting for UI changes, asserting state, and handling alerts. References apply to the latest output, so an agent should use updated references after a settled action.

Agents can also capture screenshots and video, collect logs and traces, inspect network data, gather performance samples and crash details, and profile React Native renders. Working steps can be saved as `.ad` replay scripts for local use or CI, with strict Maestro YAML export available when needed.

## Setup and configuration

Install the CLI globally with npm, then check the environment before connecting an agent:

```bash
npm install -g agent-device@latest
agent-device doctor
agent-device help workflow
```

Node.js 22.12 or newer is required. Web automation requires Node.js 24 or newer. Target-specific installation requirements apply to the platforms being automated.

An MCP client can use this configuration:

```json
{
  "mcpServers": {
    "agent-device": {
      "command": "agent-device",
      "args": ["mcp"]
    }
  }
}
```

Run `agent-device doctor` before handing the CLI to an agent. The same commands can also be used directly from a shell or through the typed `createAgentDeviceClient()` Node.js API.

## Tools and capabilities

- Inspect accessibility trees through token-efficient snapshots, refs, and selectors.
- Operate application interfaces with presses, field filling, scrolling, gestures, waits, and assertions.
- Capture screenshots, video, logs, traces, network information, performance samples, crashes, and React profiles.
- Save repeatable workflows as `.ad` scripts and export strict Maestro YAML.
- Coordinate sessions and evidence across local devices, CI jobs, and remote device environments.
- Use prompts, resources, status information, installation guidance, and version-specific help through MCP.

## Limitations and notes

Accessibility snapshots work best when applications expose clear labels, roles, and test IDs. Screenshots and video remain useful as evidence or when accessibility data is insufficient. References are short-lived and should be taken from the latest command output.

The server depends on the agent-device CLI and the platform-specific simulator, emulator, physical-device, desktop, browser, or remote-device setup required by the selected target. It is intended for coding-agent workflows and can run locally, in CI/CD, or with remote device infrastructure.

_Full upstream README: https://allmcps.com/mcp/callstackincubator-agent-device/readme_