2026.09.19
agent-device: Let AI Coding Agents Verify Running Apps
agent-device provides a CLI, MCP server, and typed Node.js API for inspecting, controlling, and verifying applications. It fits teams connecting coding agents to mobile testing, debugging, or CI, but platform coverage and environment requirements need careful validation.
Purpose
agent-device is an open-source automation and verification tool from Callstack. Its goal is to give AI coding agents a feedback loop through a running application. The project exposes a command-line interface, a built-in MCP server, and a typed Node.js API. That allows teams to use it from agents that can run a CLI, connect an MCP client, or embed it in a Node.js agent runtime.
The central workflow is inspect, act, and verify. An agent can read accessibility snapshots, target interface elements with refs or selectors, and perform presses, text entry, scrolling, gestures, waits, assertions, and alert handling. The README presents accessibility data as more useful for agent interaction than relying on screenshots alone. Screenshots and video remain available as evidence, or as a fallback when accessibility data is incomplete.
Image source: agent-device. Published by the project; the interface may vary by version.
Features and audience
The project lists iOS, Android, HarmonyOS, tvOS, Android TV, Amazon Vega OS TV, web, macOS, and Linux as targets, covering simulator, emulator, and some physical-device workflows. It also describes support for React Native, Expo, Flutter, and native applications, while noting that commands and evidence vary by target.
agent-device may suit mobile developers, end-to-end test engineers, teams asking coding agents to verify their own changes, and maintainers who want to move exploratory flows into CI. Listed capabilities include saving workflows as .ad scripts, exporting Maestro YAML, and collecting screenshots, video, logs, traces, network data, performance samples, crash details, and React profiles. The project also describes session and device ownership scoped to Git worktrees, intended to reduce interference between parallel agents, and integrations with remote device services including BrowserStack, AWS Device Farm, and Limrun.
Constraints before adoption
The platform list should not be read as equal support everywhere. The README explicitly says newer backends such as HarmonyOS and Vega OS cover only a subset of commands, and recommends checking capabilities for a specific target. Apple, Android, desktop, TV, and web paths use different underlying bridges, so snapshot, interaction, recording, and profiling behavior should be validated separately.
Environment setup is another deployment concern. The README requires Node.js 22.12 or newer, while web automation requires Node.js 24 or newer. Target-specific simulators, emulators, physical devices, drivers, permissions, or remote-cloud configuration may also be needed. The project recommends running its doctor check, but the supplied sources do not confirm any particular host setup.
Accessibility quality affects reliability. Clear labels, roles, and test IDs can make runs more stable; when an application exposes weak semantics, screenshots or video provide evidence but do not guarantee dependable automation. Teams running parallel work should also understand session scope, device claims, and stale-owner handling.
The supplied official release record identifies v0.21.4, published on September 16, 2026, with changes across Android, iOS, recording, device ownership, and runner lifecycle areas. Teams should pin a version they have validated and re-check capabilities, permissions, evidence artifacts, and CI behavior for their own targets. GitHub metadata identifies the project as MIT licensed; that does not replace review of dependencies, platform terms, or internal security procedures.
Sources
Sources checked: 2026-09-19