Drive a mobile simulator from a coding agent over MCP
A coding agent that writes mobile code cannot tell whether the code works until it can open the app and look. Limrun's MCP server gives the agent a cloud iOS simulator or Android emulator it can launch, tap through, read back, and show you live.
Add Limrun to your client
Install Limrun from your client's own directory. You sign in with your Limrun account the first time the agent uses it, and there is nothing else to configure.
- Claude (claude.ai and Claude Desktop): add Limrun from the connectors directory.
- ChatGPT: add the Limrun plugin.
Coding agents install the same thing as a plugin:
| Agent | Install |
|---|---|
| Claude Code | /plugin marketplace add limrun-inc/skills, then /plugin install limrun@limrun |
| Codex CLI | codex plugin marketplace add limrun-inc/skills, then codex plugin install limrun |
| Gemini CLI | gemini extensions install https://github.com/limrun-inc/skills |
| Cursor | git clone https://github.com/limrun-inc/skills ~/.cursor/plugins/local/limrun |
For any other MCP client, connect it to https://mcp.limrun.com/mcp. Remote MCP server covers manual setup and API-key auth for headless agents.
What MCP covers, and what the CLI covers
MCP is for working with devices: creating them, tapping, typing, reading the screen and the app's logs, and showing you the live device. Building an app and installing files from your machine take the lim CLI, because an MCP server cannot read your local files. The plugins above install both: the MCP server, plus skills that teach the agent when to switch to the CLI.
Ask for the outcome
You do not call the tools yourself. Describe what you want tested, and the agent picks the tools:
Open an iOS simulator with my app, sign in as [email protected], and tell me what the home screen shows.
Behind that request, the agent:
- Calls
create-ios-simulator, which waits until the simulator is ready and returns its instance ID. Apps can be preinstalled at creation by asset name or by public URL. - Calls
ios-screenshot-and-element-treeto read the screen. The element tree lists each element's label, accessibility ID, type, and frame, so the agent works from structure instead of guessing from pixels. - Calls
ios-usewith a batch of actions, such as tapping the email field, typing, and tapping Sign in, then reads the screen again. - Calls
ios-logsto check the app's output, because a screen that looks right can sit on top of a failed request. - Calls
delete-instancewhen it is done.
Ask it to "show me the simulator" at any point, and it calls stream-device, which shows the live, interactive device inside the chat where the client supports it. Android works the same way with create-android-emulator, android-screenshot-and-element-tree, and android-use.
Remote MCP server lists every tool and its arguments.
Troubleshooting
The agent says it cannot reach Limrun. Open your client's connector or MCP settings and finish the Limrun sign-in. In Claude Code, run /mcp, select limrun, and choose Authenticate.
The simulator disappeared in the middle of a session. Simulators stop after a period without interaction, set by your organization. Ask the agent to create the simulator with a longer inactivityTimeout, such as "30m".
The agent cannot find a row further down a list. iOS creates list rows only when they scroll into view. Ask the agent to scroll before tapping; in ios-use, scroll actions go before the tapElement in the same batch.
Taps land but the app does not respond. Ask the agent to bring the app to the front with ios-app and to check ios-logs for a crash.
The agent should not hold an organization credential. If the agent runs in a shared or untrusted sandbox, create the simulator yourself and give the agent only that one device. Sandboxed agents shows how.
Was this guide helpful?