ios-device-automation

iOS Device Automation

CRITICAL RULES — VIOLATIONS WILL BREAK THE WORKFLOW:

1. Never run midscene commands in the background. Each command must run synchronously so you can read its output (especially screenshots) before deciding the next action. Background execution breaks the screenshot-analyze-act loop.
2. Run only one midscene command at a time. Wait for the previous command to finish, read the screenshot, then decide the next action. Never chain multiple commands together.
3. Allow enough time for each command to complete. Midscene commands involve AI inference and screen interaction, which can take longer than typical shell commands. A typical command needs about 1 minute; complex act commands may need even longer.
4. Always report task results before finishing. After completing the automation task, you MUST proactively summarize the results to the user — including key data found, actions completed, screenshots taken, and any relevant findings. Never silently end after the last automation step; the user expects a complete response in a single interaction.

Automate iOS devices using npx -y @midscene/ios@1. Each CLI command maps directly to an MCP tool — you (the AI agent) act as the brain, deciding which actions to take based on screenshots.

What act Can Do

Inside a single act call on iOS, Midscene can tap, double-tap, long-press, type, clear text, scroll, drag items, zoom with two fingers, press keys, and use system navigation such as Home or the app switcher while working from the current visible screen.

Prerequisites

Midscene requires models with strong visual grounding capabilities. The following environment variables must be configured — either as system environment variables or in a .env file in the current working directory (Midscene loads .env automatically):

MIDSCENE_MODEL_API_KEY="your-api-key"
MIDSCENE_MODEL_NAME="model-name"
MIDSCENE_MODEL_BASE_URL="https://..."
MIDSCENE_MODEL_FAMILY="family-identifier"

Example: Gemini (Gemini-3-Flash)

Example: Qwen 3.5

Example: Doubao Seed 2.0 Lite

Commonly used models: Doubao Seed 2.0 Lite, Qwen 3.5, Zhipu GLM-4.6V, Gemini-3-Pro, Gemini-3-Flash.

If the model is not configured, ask the user to set it up. See Model Configuration for supported providers.

Commands

Connect to Device

If WebDriverAgent is already running and the session was created outside Midscene, pass the WDA endpoint and external session ID together:

Use the same --wda-host, --wda-port, and --session-id options on later commands when you want them to operate through that existing WDA session.

Launch an App, URL, or Deep Link

Use the built-in launch capability when you want to start from a known app or route before the rest of the task. Give it the most specific target you have, such as a bundle ID, web URL, deep link, or phone/mail link. Typical targets include com.apple.Preferences, https://www.apple.com, myapp://profile/user/123, and tel:+1234567890.

Send a Direct Device Request

Use this when the task needs lower-level device control instead of a normal visible UI interaction:

This does not run an ADB command. On iOS, the underlying operation is an HTTP request to WebDriverAgent, typically GET http://<wdaHost>:<wdaPort>/session/<sessionId>/wda/screen.

Take Screenshot

After taking a screenshot, read the saved image file to understand the current screen state before deciding the next action.

Perform Action

Use act to interact with the device and get the result. It autonomously handles all UI interactions internally — tapping, typing, scrolling, swiping, waiting, and navigating — so you should give it complex, high-level tasks as a whole rather than breaking them into small steps. Describe what you want to do and the desired effect in natural language:

Assert Current Screen State

Use assert to verify that the current screen satisfies a natural language condition. It does not perform UI actions; it checks the visible screen state and passes only when the assertion is true. Use this for validation, QA checks, and final state verification after act.

By default a failed assertion throws an AI-generated reason. Pass --message to throw a custom error message instead, which is useful for surfacing the intended outcome in QA and CI logs.

When the assertion needs to compare against a reference image (icon, logo, screenshot), pass --image for the URL/path and --image-name for its display name. Each --image may be an http(s) link, a data: URI, or a local file path. Repeat both flags in matching order when you need to attach more than one image. Add --convertHttpImage2Base64 true when the model cannot reach the URL directly. Requires @midscene/ios@1.9.0+.

Record and Assert Transient UI

Use a recording when the state to verify may disappear before a current-screen assertion runs, such as a toast, loading banner, animation, or transition:

Pass target flags, capture flags, and --output to record start, then wait for Recording. Press Ctrl+C to stop and save. Keep the recorder as a foreground process in its dedicated terminal; never add shell &. Perform the interaction manually or from a second terminal, send Ctrl+C to the recorder, and wait until it prints the saved path before asserting. Recording uses the WDA MJPEG frame source when enabled and otherwise falls back to periodic screenshots. Optional capture flags are --interval-ms, --max-frames, and --watchdog-ms; --max-frames caps sampled frames, and the manifest may contain one additional final representative frame. The default watchdog finalizes and saves the recording after five minutes, while --watchdog-ms 0 disables that safety limit. The output is a JSON manifest plus an adjacent <name>.frames image directory, not an encoded video or archive. The manifest contains relative JPEG/PNG paths and no base64 image bodies. Keep or move the JSON file and image directory together, and pass the JSON path to assert --record. Use ordinary assert without --record when only the current screen matters.

Use a Reference Image for Precise Targeting

When the user provides a screenshot, icon, logo, or reference image and wants an exact visual match, prefer tap --locate instead of a generic act --prompt. Pass --locate as JSON. The prompt describes the target, images supplies named reference images, and convertHttpImage2Base64: true is useful when the image URL may not be directly accessible to the model.

The same locate JSON shape also works for other commands that accept a locate parameter.

Disconnect

Consume Report Files

The generated HTML report is recommended for human reading first. It includes step-by-step execution details and replay videos for each operation, which makes it much easier to understand what happened and troubleshoot problems.

If another skill or tool needs to consume the report, first convert it with report-tool from the same platform CLI package. Prefer Markdown for LLM-based workflows. Use JSON when the report needs to be processed programmatically.

Workflow Pattern

Since CLI commands are stateless between invocations, follow this pattern:

1. Connect to establish a session
2. Launch the target app and take screenshot to see the current state, make sure the app is launched and visible on the screen.
3. Execute action using act to perform the desired action or target-driven instructions. Use assert for the resulting screen state, or keep record start --output ... running in a dedicated terminal during transient-state workflows, stop it with Ctrl+C, and then use assert --record.
4. Disconnect when done
5. Report results — summarize what was accomplished, present key findings and data extracted during the task, and list any generated files (screenshots, logs, etc.) with their paths

Best Practices

1. Be specific about UI elements: Instead of vague descriptions, provide clear, specific details. Say "the Settings icon in the top-right corner" instead of "the icon".
2. Describe locations when possible: Help target elements by describing their position (e.g., "the search icon at the top right", "the third item in the list").
3. Never run in background: Every midscene command must run synchronously — background execution breaks the screenshot-analyze-act loop.
4. Batch related operations into a single act command: When performing consecutive operations within the same app, combine them into one act prompt instead of splitting them into separate commands. For example, "open Settings, tap Wi-Fi, and check the connected network" should be a single act call, not three. This reduces round-trips, avoids unnecessary screenshot-analyze cycles, and is significantly faster.
5. Choose the right verification window: Use assert --prompt "..." for the current screen. For a toast, banner, animation, or transition, run record start --output ... in a dedicated terminal, perform the interaction, stop recording with Ctrl+C, wait for the saved-path message, then pass the artifact to assert --record.
6. Always report results after completion: After finishing the automation task, you MUST proactively present the results to the user without waiting for them to ask. This includes: (1) the answer to the user's original question or the outcome of the requested task, (2) key data extracted or observed during execution, (3) screenshots and other generated files with their paths, (4) a brief summary of steps taken. Do NOT silently finish after the last automation command — the user expects complete results in a single interaction.
7. Prefer tap --locate when a reference image is provided: If the user shares a screenshot, icon, or logo and wants that exact visual target, use tap --locate with a multimodal locate JSON object such as { "prompt": "...", "images": [...] } instead of relying only on act --prompt.

Example — Alert dialog interaction:

Example — Form interaction:

Improve Precision (Deep Locate / Deep Think)

Two optional global flags help when Midscene struggles with a task. Put them anywhere in the command (before or after the sub-command); once set, the relevant operations use them by default, so you don't pass a per-call parameter.

• --deep-locate — spends an extra round of visual reasoning to pinpoint the target element. Use it when an action interacts with the wrong spot (location drift / offset). It applies to every operation that locates an element, including tap --locate and the locating that happens inside act.
• --deep-think — plans act with deeper reasoning (richer context and sub-goal decomposition). Use it for complex, multi-step act instructions; it only affects planning.

Both trade a little speed for better results, and you can combine them.

Troubleshooting

WebDriverAgent Not Running

Symptom: Connection refused or timeout errors. Solution:

• Ensure WebDriverAgent is installed and running on the device.
• If another tool created the WDA session, pass that session explicitly with --session-id plus the matching --wda-host and --wda-port.
• See https://midscenejs.com/usage-ios.html for setup instructions.

Device Not Found

Symptom: No device detected or connection errors. Solution:

• Ensure the device is connected via USB and trusted.

API Key Issues

Symptom: Authentication or model errors. Solution:

• Check .env file contains MIDSCENE_MODEL_API_KEY=<your-key>.
• See https://midscenejs.com/zh/model-common-config.html for details.

@midscene/* Dependency Version Outdated

Symptom: Unexpected behavior, missing features, or version mismatch errors. Solution:

• Check local versions: npm ls @midscene/ios @midscene/core @midscene/shared (or pnpm why @midscene/ios).
• Check latest versions: npm view @midscene/ios version, npm view @midscene/core version, npm view @midscene/shared version.
• Upgrade dependencies: npm i @midscene/ios@latest @midscene/core@latest @midscene/shared@latest.