On September 14, the consumer version of the Doubao Phone Assistant launched, with its first device — the Nubia NaviX Ultra — going on sale September 16. Released the same day, the SAEP protocol gave third-party apps their first formal way to say "no" to AI — ByteDance seized the initiative in drafting the rules for "whether AI can tap your app." The era of "the user decides" has been broken into a three-party contract.
Hitching a ride on Nubia, it finally gets a real mass-produced device
Doubao Phone Assistant no longer lives only in demo videos — its first mass-produced home is a Nubia phone.
From preview to consumer version, Doubao has stopped talking concepts and started talking stability and everyday usability. The first device to carry it is the Nubia NaviX Ultra, and pre-orders are now open.
The hardware highlights are dense: a 200-megapixel main camera, a 7.62mm-thin body packing a 7100mAh battery, and the "world's first AI agent phone" label already slapped on. Doubao Assistant partnered with the phone maker at the system level. AI can be woken via voice or a dedicated AI key, and the AI key comes with fingerprint authentication(verifies it's the phone's actual owner when pressed) — once the owner is confirmed, the assistant can be summoned without a second unlock.
The new version adds one capability: on-screen Q&A — no screenshots, no app switching. The assistant reads the screen directly and continues handling the task. There are three ways to manually save a memory: voice input, a three-finger swipe up, or pressing the AI key and volume key together.
The most-watched "operating the phone" feature hasn't been pulled — it's been opened up in beta form, wrapped in a new protocol Doubao calls SAEP that lets third-party apps declare what AI assistants may and may not do inside them.
Two ways to wake it, one fingerprint press hands over the screen
Doubao Phone Assistant offers two wake methods: say a word, or press the AI key. A fingerprint module is embedded in the key — after verifying the owner, it launches the assistant directly, skipping one unlock. Voice wake takes a different path — far-field pickup plus internal noise suppression, designed for far-field and noisy environments.
On-screen Q&A is the change closest to daily life. Open the camera and point it at a room, then say "based on this décor style, help me pick a cabinet under 1.2 meters and under 1,000 yuan." The assistant combines what's on screen and returns e-commerce results. No screenshots, no app switching — the glue between the camera and the shopping app is applied by the assistant itself.
Local search has also been wired into this chain. Ask "when does my Hong Kong-Macau travel permit expire," and the assistant digs through your photo album, messages, and notes on its own. The recording feature has been upgraded too — the official integration brings in Feishu Minutes(the meeting transcription and summary tool built into ByteDance's office app "Feishu"), so recorded content can continue to be searched locally — e.g., follow up with "what product suggestions came up in yesterday's user interview." Three save entry points are offered: say a voice command, swipe up with three fingers, or press the AI key and volume key together — the current screen gets pinned into the local memory bank.
The official positioning of the consumer version is a stabilized, everyday-ready version of the preview. Released alongside it is the Screen Automation Operation Declaration Protocol (SAEP), which hands the "can AI tap you" switch to the apps themselves. How the mechanism lands in practice, and what SAEP grows into after its 30-day public comment period — that's broken down next.