Let Codex Verify Your Mobile App on a Live Device

How it works

1. 1. Install the CLI and log in yourself

   Do the browser login yourself, once. After that, Codex can use the CLI on its own. For unattended runs, an API token in the environment works instead.

   npm install -g vibeview
   vibeview login

2. 2. Let Codex’s commands reach the network

   Codex’s default workspace-write sandbox keeps network access off for the commands it runs, and every vibeview command talks to VibeView, so they fail until you turn network access on in ~/.codex/config.toml (or approve each command when Codex asks).

   [sandbox_workspace_write]
   network_access = true

3. 3. Install the skill and the AGENTS.md line

   The vibeview-agent skill is plain instructions: which build to upload, how to start a session, to read the tree before acting, and to stop the session at the end. The skills CLI installs it for Codex from the public mirror. vibeview agent-setup also prints a one-line instruction to add to AGENTS.md, which Codex reads, so it verifies UI changes on a device without being reminded.

   npx skills add vibeview/skills

4. 4. Or register the MCP server

   Over MCP, Codex gets the same commands as tools, named after the CLI verbs, with underscores for hyphens (ui_tree, tap, type, screenshot, logs), plus upload_app, dev_start and dev_stop so the whole loop needs no shell at all. One command adds the server; the config.toml version is below.

   codex mcp add vibeview -- vibeview mcp
   
   # or in ~/.codex/config.toml
   [mcp_servers.vibeview]
   command = "vibeview"
   args = ["mcp"]

5. 5. Ask for the change and the check

   Put “and verify it on the device” in the task. Codex uploads the debug build (or a Roku channel .zip), which it reuses until a native change, then starts a detached session and works from the UI tree: every tap names an element reference, not a guessed coordinate. A React Native session tunnels your local Metro bundler, so Codex sees an edit on the device about a second after saving it. On a TV, the press command drives the d-pad and the focus command walks to an element by name. When a screen needs a real login or a risky confirmation, Codex stops and asks you.

   vibeview upload-app ./android/app/build/outputs/apk/debug/app-debug.apk
   vibeview dev --detach --json
   vibeview ui-tree --session <id>
   vibeview tap @e5 --session <id>
   vibeview screenshot --out ./check.png --session <id>
   vibeview dev-stop

Full setup instructions live in the documentation.