vibeview
Official# VibeView agent skills
[](https://www.npmjs.com/package/vibeview)
[](https://github.com/mcp/vibeview/vibeview)
Skills and an MCP server that let a coding agent verify its work on a **live cloud device**: upload a build, start an iOS, Android, Apple TV, Android TV or Roku session, drive the app by UI-tree refs, read logs, take screenshots, and stop when only a human can act.
[VibeView](https://vibeview.io) streams iOS simulators, Android emulators, Apple TV and Android TV simulators, and Roku devices to a browser and to agents. No Mac, no local simulator, no device on the desk. Docs: [AI agent control](https://vibeview.io/docs/agent-control).
<p align="center">
<img src="assets/agent-session.png" alt="Claude Code using the VibeView MCP server: it starts an iPhone session, reads the UI tree, taps a card by its ref, takes a screenshot and stops the session" width="800">
</p>
## What your agent can do
- **Check its own change in your React Native app.** It starts `vibeview dev` against your local Metro, so every edit hot-reloads on the cloud device, then it looks at the screen and the logs to see whether the change works.
- **Build the app in the cloud and run that build.** React Native, Expo, Flutter, Capacitor, Kotlin Multiplatform or native iOS and Android: with a shell, it runs `vibeview build --cloud` (no Mac needed for iOS), then starts a device on the finished build.
- **Drive the app like a person.** It reads the screen as a UI tree with short refs (`@e7`), taps, types and swipes by ref, presses TV remote buttons, opens deep links, and takes screenshots.
- **Write and run tests.** It saves plain-language steps as a test, then runs one test or a whole group and reads which step failed and why. AI steps are in beta.
- **Hand off to you** when a step needs a person, such as a login or a payment, with a link to the live session.
It works on iPhone, iPad, Android phones and tablets, Apple TV, Android TV and Roku. You can watch every session live in the dashboard while the agent works.
<p align="center">
<img src="assets/device-iphone-duo.png" alt="An iPhone Duo session in the VibeView dashboard, half folded, with the UI tree panel listing elements by ref (@e0, @e1, …)" width="800">
</p>
<p align="center">
<img src="assets/device-appletv.png" alt="An Apple TV session in the VibeView dashboard" width="49%">
<img src="assets/device-roku.png" alt="A Roku session in the VibeView dashboard" width="49%">
</p>
## Install
You need the `vibeview` CLI, logged in:
```bash
npm install -g vibeview
vibeview login
```
Then pick whichever fits your agent.
**Any agent that reads skills** (Claude Code, Cursor, Codex, and others via the `skills` CLI):
```bash
npx skills add vibeview/skills
```
**Claude Code, as a plugin** (installs the skill and registers the MCP server in one step):
```text
/plugin marketplace add vibeview/skills
/plugin install vibeview@vibeview
```
**Claude Code, from the CLI you already installed** (copies the skill into `~/.claude/skills/` and offers to register the MCP server):
```bash
vibeview agent-setup
```
**Any MCP client**: run `vibeview mcp` over stdio.
```json
{
"mcpServers": {
"vibeview": { "command": "vibeview", "args": ["mcp"] }
}
}
```
With Claude Code specifically: `claude mcp add vibeview -- vibeview mcp`.
## What's here
| Path | What it is |
| --- | --- |
| `skills/vibeview-agent/SKILL.md` | The agent workflow: build requirements, the dev loop, verifying with `ui-tree`, screenshots and logs, TV d-pad navigation, the human hand-off rule, cleanup, and the full command reference for both the CLI and MCP. |
| `.claude-plugin/` | Claude Code plugin manifest: the skill plus the `vibeview mcp` server. |
| `.mcp.json` | The same MCP server in the standard root-level form that plugin and MCP directories scan for. |
## Where this comes from
The canonical copy of the skill ships inside the `vibeview` npm package (`vibeview agent-setup` copies it from there). This repository mirrors it so agents and registries can discover it on GitHub. Versions track the CLI.
Issues and questions: support@vibeview.io.
TDQS
Scored across 38 tools
Most tools target a clearly distinct action+object: tap vs long_press vs drag vs type vs clear_text are unambiguous, and the descriptions actively separate the easily-confused trio swipe/scroll/scroll_to and find vs ui_tree. The remaining soft spots are dev_stop vs stop_session (overlap is explained in the text but still requires reading) and scroll vs swipe, which an agent could mix up without the descriptions. tap_focused vs focus vs tap on TV is handled well by explicit 'activates'/'without activating' wording.
Everything is lowercase snake_case with a consistent terse style, and the namespaced families (dev_*, list_*, *_test, run_*/get_run, set_*) are predictable. Deviations are minor: some names are bare verbs (tap, press, wait, find, focus, rotate, shake, alert) while others are verb_noun (open_url, clear_text, scroll_to, relaunch_app), and a few are noun-first (ui_tree, logs, screenshot). Still readable and guessable.
38 tools is a heavy surface, but the breadth is real: device interaction, device/system controls, session lifecycle, app/build catalogs, and the test runner are genuinely separate workflows. Individual tools each map to a distinct capability rather than padding, yet the set would benefit from tighter grouping or fewer near-duplicate scroll variants. Borderline over-scoped for a single MCP server.
Device control and inspection are well covered (input, gestures, rotation, posture, locale, location, logs, screenshot, tree, session start/stop/reload, app and build listing). The obvious gap is on the test surface: create_test/run_test/run_group/get_run exist but there is no list_tests, list_groups, update/delete_test, or run history listing, so an agent cannot discover existing tests or clean up. No session listing either — stop_session wants an id the server never helps you find.