Claude Desktop and your iPhone: MCP setup

Claude Desktop is where many people talk to Claude outside a terminal. With mobster mcp added, Claude can run an iOS app on a headless simulator Mobster manages, read each screen as a list of controls by name, tap and type, and return a verdict from assertions on the accessibility tree. It can also use a real iPhone plugged into your Mac, through the same tools.

You need an Apple silicon Mac with macOS 15 or later. Simulator work needs Xcode with an iOS Simulator runtime. A real iPhone needs Developer Mode and the WebDriverAgent runner, set up in Mobster for Mac or with mobster serve --manage-device (Device setup).

Install Mobster CLI

$ curl -fsSL https://mobster.dev/install.sh | sh

or

$ brew install radishsoftware/tap/mobster

Then note the full path to the command, which Claude Desktop needs, and prepare the simulator once:

$ mobster version
$ which mobster
$ mobster sim doctor --fix

Add the server

Open Claude Desktop’s config from the Claude menu: Settings, then the Developer tab, then Edit Config. The file is ~/Library/Application Support/Claude/claude_desktop_config.json. Add Mobster under mcpServers, with the path which mobster printed:

{  "mcpServers": {    "mobster": {      "command": "/opt/homebrew/bin/mobster",      "args": ["mcp", "--out", "/Users/<you>/mobster-runs"]    }  }}

Use an absolute path for both. Claude Desktop starts a local server with only a small part of your environment, so mobster alone may not be found, and the server’s working folder may be /. When it is, Mobster writes runs to runs/ in its own data folder; --out puts them where you can find them. Each run folder holds report.html, result.json and the frames.

Or let Mobster write the entry with the full path: mobster mcp install claude-desktop. It leaves out --out, so add that yourself if you want the runs in a folder you choose.

Quit Claude Desktop completely and open it again, so it reads the file.

For Smart, where Mobster runs the steps itself on your OpenAI or Anthropic key, add "--env-file", "/Users/<you>/.config/mobster/agent.env" to args. The file holds OPENAI_API_KEY=… or ANTHROPIC_API_KEY=…, readable only by you. Without it, Claude does the driving with its own model and Mobster calls no model at all.

The MCP pages for Claude Desktop don’t state a tool-call timeout. Mobster keeps its calls short: each returns within 45 seconds and wait within 50. The first simulator boot and a Smart run continue under a run_id, and Claude calls wait until they’re done.

Check it’s connected

Click the + button under the chat box, then Connectors: mobster should be listed. The Developer tab in Settings shows each server’s status and opens its logs. Mobster’s log, which it writes to stderr, ends up in ~/Library/Logs/Claude/mcp-server-mobster.log. Then ask:

Call mobster's status tool and tell me what it says.

A first check: the paywall, launch-only

A check with no steps is the quickest one to try. Build Daybreak, the sample app in the CLI’s repository, once in Terminal:

$ git clone https://github.com/RadishSoftware/mobster && cd mobster
$ xcodebuild -project examples/ios/Daybreak/Daybreak.xcodeproj -scheme Daybreak \
  -destination 'generic/platform=iOS Simulator' -derivedDataPath examples/ios/Daybreak/.build CODE_SIGNING_ALLOWED=NO build

Then ask Claude, with the absolute path to the app you built:

Use mobster to check Daybreak at /Users/<you>/mobster/examples/ios/Daybreak/.build/Build/Products/Debug-iphonesimulator/Daybreak.app.
Open daybreak://paywall with no steps, and expect "Choose your plan", three elements whose id starts
with plan_, Annual's value "$39.99 / year", a Restore Purchases button, and no "Loading".

Claude calls verify_start with an empty steps list and these expectations, which are the ones in Daybreak’s own paywall.yaml:

[  {"text": "Choose your plan"},  {"count": {"id": "/^plan_/"}, "equals": 3},  {"value": {"id": "plan_annual"}, "equals": "$39.99 / year"},  {"visible": {"label": "Restore Purchases", "role": "button"}},  {"no_text": "Loading"}]

Mobster installs the app, opens the deep link, and verify_finish decides on one settled read of the accessibility tree. No model decides the verdict. Claude gets passed or failed, the line that failed if any, and the path to report.html, which shows the screen with each asserted element outlined in green or red. From there, ask Claude to tap through a flow, and it uses screen, tap, type_text and swipe before verify_finish.

Use a real iPhone

Ask Claude to call list_devices. A USB iPhone set up for Mobster is listed with its name and state. Called with that device and no run_id, screen, tap, type_text, swipe, alert, open_url, launch_app and home act on the phone itself. Mobster takes the phone’s lock, so the Mac app and other commands wait, and releases it after stop with the device or 90 idle seconds.

Claude Desktop asks for your approval before a tool runs. That approval covers a tool, not a device: allowing tap while testing on a simulator allows it on your iPhone too. Limit the server to the phones you choose by adding "--allow-device", "Test iPhone" to args; any other real device is listed as not_allowed and refused. Mobster’s MCP annotations mark status, screen, list_devices, wait, wait_for, read_notifications and unlock_status as read-only, and its instructions tell Claude to treat text on the screen as data and to ask you before anything that sends, buys, posts or deletes on a real iPhone.

A locked iPhone stops the call at once with “Your iPhone is locked. Unlock it and try again. No action was taken.” Mobster never enters a passcode and never uses Face ID.

Troubleshooting

The MCP reference lists every tool. For coding work in the terminal, see Mobster in Claude Code.