DeepSeek Harness (DSH) is DeepSeek's open-source agent framework, similar to Claude Code but powered by DeepSeek's models. Its plugin ecosystem has been growing rapidly recently. The team led by Fini.Yang has released its fourth plugin, dsh-ios, which enables automation for both iOS simulators and physical iOS devices.

This plugin allows Agents within DSH to directly view and control iOS simulators and physical devices. It supports building and running apps, performing taps, swipes, text input, checking running processes, call stacks and memory leaks. It comes with 21 built-in Agent tools. Real-time device previews are displayed in the sidebar, supports human-AI collaborative operation, and also enables hot reload for SwiftUI previews.

The first half of the demo video shows the simulator in use, and the second half shows operation on a physical device.

Installation is straightforward:

```

dsh plugin --profile web add @zseven-w/dsh-ios@latest

dsh web

```

Instead of guessing coordinates to interact with UI elements, the Agent taps controls via their identifiers through an AXe-assisted UI tree, and uses Vision OCR to locate elements by on-screen text when the UI tree is empty. For list/feed applications, there are dedicated tools for row reading and row tapping, which verify operation effectiveness by parsing counters. Support for physical devices is implemented via WebDriverAgent: connect your iPhone via USB, build and launch WDA, and control signals + screen feeds are transmitted through a loopback tunnel. All visual data is routed to the browser via signed endpoints of the DSH web server; the browser never accesses the simulator port directly, and authentication tokens expire after 10 minutes.


When a netizen asked about token consumption in the comments, the author replied: "For 10 or fewer consecutive operations, it's a few hundred KB. We use flash [model] because it's fast, right?" Running consecutive operations with cheap and fast models still incurs noticeable costs. This is roughly the current reality of Agents operating physical devices.

Before dsh-ios, the same team had released the third plugin, dsh-crew, which allows Claude Code / Codex to schedule DSH Agents as native sub-agents, with status, progress and token consumption all visible in the host UI. DSH can also in turn call the multimodal and image generation capabilities of Claude / Codex / Grok / Antigravity. In the screenshot of dsh-crew, 18 worker agents are translating a README in parallel, the host side displays native sub-agent progress, and the DSH side shows the source, slot and token consumption of each task.



Along with the previously released OpenPencil (design document inspection and editing) and Noema (long-term memory), DSH's plugin ecosystem is now taking shape.

While many are still debating whether AI can write code, AI has already learned to open simulators and interact with apps on its own. The imagination for this toolchain has been unlocked, and it's now up to developers to explore its potential.