callstack/agent-device

Mobile app automation and verification for AI coding agents. CLI, MCP server, and typed Node.js API for iOS, Android, HarmonyOS, TV, web, macOS, and Linux.

What it solves

agent-device provides a feedback loop for AI coding agents, allowing them to verify their code changes in a live running application. It eliminates the need for agents to rely solely on screenshots, providing them with structured data to inspect, control, and debug apps across multiple platforms.

How it works

The tool operates as a CLI, an MCP server, or a Node.js API. It uses a local accessibility bridge for iOS Simulators, XCTest for physical iOS/tvOS, ADB for Android, and various other platform-specific helpers (HDC for HarmonyOS, Vega CLI for Vega OS, AT-SPI for Linux).

Agents interact with the app using token-efficient accessibility snapshots and references (refs) to elements. The system manages device ownership to prevent parallel agents from conflicting on the same simulator or emulator, and it can connect to remote device clouds like BrowserStack and AWS Device Farm.

Who it’s for

It is designed for AI coding agents (such as Claude Code, Cursor, Windsurf, and Cline) and developers building their own agents using the AI SDK or Eve, who need to automate the verification of mobile, TV, and desktop applications.

Highlights

  • Multi-platform support: Works across iOS, Android, HarmonyOS, tvOS, Android TV, Amazon Vega OS, web, macOS, and Linux.
  • MCP Server: Built-in Model Context Protocol (MCP) server for easy integration with compatible AI agents.
  • Evidence Capture: Ability to capture screenshots, video, logs, traces, and performance samples as evidence for review.
  • CI/CD Integration: Supports saving workflows as .ad scripts for automated replay in CI environments.
  • Device Coordination: Coordinates device access across parallel agent worktrees to prevent conflicts.

Related

  • Project
    lycorp-jp/sim-use

    A cross-platform CLI that enables AI agents to observe and act on iOS and Android screens using token-efficient UI outlines and alias-based interactions.

  • Project
    AlexGladkov/claude-in-mobile

    `mcp-devices` (aka `claude‑in‑mobile`) is a modular Node/Rust server that lets Claude‑style AI agents control Android, iOS, web, desktop, Aurora, and HarmonyOS devices. Install the base (`npm i -g mcp-devices`), add platform plugins (e.g., `@mcp-devices/plugin-android`), enable them, and point an MCP‑compatible client at the server. The tool provides actions like screenshots, taps, UI inspection, live debugging, and testing, enabling AI‑driven device automation and embodied‑agent research.

  • Project
    openchamber/openchamber

    OpenChamber is an open‑source, cross‑platform workspace that lets developers start AI coding agents (via OpenCode), monitor their progress, compare multiple model runs, review changes with guided walkthroughs, preview the app being edited, and integrate directly with GitHub—all while syncing sessions across desktop, web/PWA, VS Code, and mobile devices.

  • Project
    ai-dashboad/flutter-skill

    An AI-native E2E testing tool that connects AI agents to apps via MCP, allowing them to test and explore software across 10 platforms using natural language instead of test code.

  • Project
    Kritt-ai/open-kritt

    An open-source security research platform that orchestrates AI agents to find, validate, and rank vulnerabilities in code through focused, parallelized workflows.