What does this skill do, and when should you use it?
agent-device is a device-interaction skill for Apple-platform apps and Android devices. It covers iOS, tvOS, macOS, and Android, with workflows for app navigation, snapshots, screenshots, taps, text entry, scrolling, and UI information extraction. Normal device tasks must begin by loading bootstrap-install.md and exploration.md. The skill favors read-only inspection first and verification after meaningful UI changes.
It routes device work through setup and exploration workflows, confirms the target, app installation, and open app session, then uses operations such as snapshot, snapshot -i, get, is, find, fill, type, and back to inspect or interact with the UI. It can re-snapshot after changes, capture evidence when needed, and close the session.
- A React Native developer needs to verify visible text or UI structure in an iOS or Android app.
- A mobile QA engineer needs to tap, type, scroll, and verify an already-open app flow.
- An AI coding assistant user needs to capture screenshots or extract UI information from a device session.
- An engineer debugging an iOS, tvOS, macOS, or Android app needs to inspect its state on the target device.
How do you install this skill?
- The skill can control real devices or simulators and may install, relaunch, record, and collect logs or network data; confirm target, permissions, session scope, and sensitive-data handling before execution.
- Remote tenancy examples involve daemon URLs, authentication tokens, and lease identifiers; do not copy real credentials or log payloads into shared environments.
- There is no evidence for mainland-China network reachability, Chinese-language output, or agent-device version compatibility.
- Shell / CLI
- Local filesystem
The repository README says Codex can add its bundled plugins through .agents/plugins/marketplace.json and that the repository can also be cloned for skill installation. It does not provide a standalone installation command for agent-device; the skill is located at plugins/vendored/.agents/skills/agent-device/. The repository is licensed under MIT.
tmp="$(mktemp -d)"
git clone --depth 1 https://github.com/callstackincubator/agent-skills.git "$tmp"
mkdir -p ~/.claude/skills
cp -R "$tmp/plugins/vendored/.agents/skills/agent-device" ~/.claude/skills/
rm -rf "$tmp"Generated from the source repository and skill path; it copies only this skill's folder. If the author's install steps above differ, follow those first. To scope it to one project, replace ~/.claude/skills with that project's .claude/skills.
How do you use this skill?
Once installed, send your agent any of these to trigger it:
- Use agent-device to open the current app, check whether the login button is visible, and capture evidence if needed.
In an AI coding assistant that can access the installed skill, give it a concrete device task such as: “Use agent-device to open the current app, check whether the login button is visible, and capture evidence if needed.” Before a normal device task, load references/bootstrap-install.md and then references/exploration.md. Pin the correct target and open a stable app session before inspecting or interacting.
What are this skill's strengths and limitations?
- Supports iOS, tvOS, macOS, and Android.
- Combines UI inspection, interaction, screenshots, and UI information extraction.
- Starts with conservative, read-only inspection by default.
- Defines clear setup, exploration, and verification routing.
- Requires the correct target, installed app, and open app session.
- Normal tasks depend on two required reference files whose contents are not included here.
- The source does not document detailed simulator or physical-device setup.
- The source provides no standalone test suite or platform-coverage evidence for this skill.
How does this skill compare with similar options?
Side by side with related skills; every score comes from the same FSRS standard.
| Skill | FS score | Stars | Last updated | License |
|---|---|---|---|---|
| Cross-Platform Device Operator this page | 58 · Recommended | ★ 1.7k | 18d ago | MIT |
| Playwright Browser Automation Skill | 68 · Recommended | ★ 3.2k | 2mo ago | MIT |
| playwright-cli Browser Automation Skill | 56 · Use with care | ★ 11k | 29d ago | MIT |
| Playwright Browser Automation Skill ✓ OpenAI · Official | 52 · Use with care | ★ 28k | 4mo ago | — |
| Web App Testing Skill | 49 · Use with care | ★ 3.1k | 8mo ago | MIT |
How did FollowSkills review this skill?
The documentation emphasizes read-before-write behavior, smallest reversible actions, target pinning, and protection of sensitive tokens and logs. Points are deducted because installation, app launching, UI mutations, and remote HTTP operations lack explicit per-action user confirmation, a general rollback path, and complete data-flow disclosure; publisher identity is unverified.
Layered references cover bootstrap, exploration, debugging, verification, macOS, and remote tenancy, with failure categories, retry guidance, and stop conditions. Points are deducted because there are no committed tests or execution evidence for key paths, and availability of agent-device, platform tooling, signing, and remote daemons is not established.
The audience, supported platforms, common scenarios, trigger commands, and boundaries between read-only inspection and interaction are fairly clear. Points are deducted because Chinese-language support is not described, mainland-China reachability is unproven, and boundaries for nonstandard UI or missing accessibility trees remain limited.
The main file acts as a router with progressive references and includes setup, sessions, selectors, troubleshooting, and examples; repository context provides an MIT license and scheduled vendored updates. Points are deducted because the selected skill has no version, changelog, explicit maintainer responsibility, or clear update path, and dependency/CLI version assumptions are not centralized.
The command flows and QA loops are directly actionable for device discovery, session setup, UI interaction, and evidence capture, with clear core goals. Static calibration limits this to 7; points are deducted because execution was not performed and correctness, cross-platform consistency, and comparative benefit over manual workflows cannot be confirmed from files alone.
The source is auditable and provides concrete commands, parameters, error categories, and failure semantics. Points are deducted because there is no committed test suite, CI coverage, or third-party execution evidence; claims about CLI behavior remain primarily author documentation without corroboration.
Open a dimension to read why it scored that way
Evidence confidence：Low — Mostly static review, author material or a limited demo; useful for discovery, not high-risk decisions.
See the full review method →