← Back to stories A technician examines a circuit board under a microscope for electronic repair.
Photo by Tima Miroshnichenko on Pexels
钛媒体 2026-03-18

They’ve built an “unblockable” Doubao Phone (豆包手机) and secured tens of millions of yuan in angel funding

A new take on cross‑app automation

Two former hardware engineers, Zhang Zhiyong (张志勇) and Shan Wenbang (单文榜), say they have built what they call an "unblockable" Doubao Phone (豆包手机)‑style agent by shipping ZeroFlow — a multimodal agent from ZeroFlow (依零科技) that uses Android's Accessibility Service to perform cross‑platform automation. How does it differ from earlier attempts like Doubao Phone or Zhipu AI's (智谱) AutoGLM? The key, the founders argue, is reliance on a system accessibility layer that lets an agent "see" and interact with UI elements (click, swipe, type) in a way app vendors cannot easily lock down without breaking their own user experience.

Technical tradeoffs and engineering pain points

ZeroFlow reportedly reads screen content, identifies buttons and inputs, and executes gestures, using a domestic multimodal model stack tuned for low token cost. That avoids the high‑privilege partnerships Doubao relied on and the remote‑VM adb workaround used by some projects. But the founders emphasize nontrivial engineering work: anti‑automation web layouts, hidden DOMs and noisy popups make element‑level understanding hard. The team says the answer is a mix of small, cheap models to prefilter screenshots and larger models for intent resolution — an approach that shrinks average prompt length by about 40% and lowers token costs by an estimated ~30% for typical users.

Safety, localization and geopolitical context

ZeroFlow (依零科技) has explicitly designed sandboxing and small‑model desensitization to limit exposure of secrets — workspace isolation hides keys and a monitoring model redacts sensitive interactions before they hit the large model. The choice to optimize for domestic models such as Kimi and DeepSeek reflects both engineering realities and geopolitical pressures: OpenClaw (小龙虾) and other projects built on OpenAI/Anthropic Tool Calling face adaptation problems inside China, and access to Western LLM tool chains can be constrained by policy and commercial barriers. Is local adaptation a necessity or an advantage? The team argues it’s both.

Funding and next steps

It has been reported that ZeroFlow (依零科技) has raised a near‑RMB‑10 million angel round from individual angels and Shangshi Capital (尚势资本) to polish product features and push distribution. The founders pitch ZeroFlow not as a niche developer tool but as a general‑purpose agent for knowledge work — finance, operations, content and beyond — with a browser‑first, zero‑install user experience: open a site, sign up, start automating. Whether app vendors, regulators or security researchers will accept an agent that claims to be "unblockable" remains the open question.

AISmartphonesSpace
View original source →