Open-Source WeChatBridge Hits iOS: Pipe WeChat Chats to AI Agents and Obsidian Straight From Your Phone
WeChat's chat records rarely leave the phone; this app changes that with a share-sheet flow that pipes them to AI agents and Obsidian without touching a desktop. The development method is the other half: acceptance-spec-driven agent work compressed 18 working days into six hours and is directly transplantable to any long-horizon coding task.
WeChatBridge, the open-source tool that forwards WeChat chat logs to AI agents and Obsidian, now runs on iOS. The new app works from the system share sheet: pick messages in WeChat, forward them to WeChatBridge, and they land as cards in an on-device inbox. From there you can send them to ChatGPT, Doubao, Qwen, Claude, DeepSeek, or Obsidian, with scene-specific prompts that auto-copy. Over 100 messages are handled by batch collection with dedup. Everything stays on device; no server upload. The app is in App Store review, with early access through the GitHub community.
The development story is just as notable. The team gave Kimi K3.1 Pro, a Chinese open-source model, a detailed spec and acceptance criteria, then let it run the whole 18-working-day plan in about six hours. The key was not describing requirements but defining "done": each milestone had exit gates with explicit simulator and real-device tests, and the agent self-checked its own output, finding and fixing five bugs during acceptance runs. The lesson the author draws: "Don't chat about requirements with it; define acceptance with it."
The acceptance-spec method is the real differentiator: the model ran 18 days of planned work in six hours because every milestone had exit gates, and unverifiable items were explicitly marked "pending real device" instead of being waved through. That discipline, not the model's raw power, is what made the agent self-correcting.
Kimi K3.1 Pro's performance on a multi-hour autonomous coding run, including finding and fixing its own bugs, suggests open-source Chinese models are now competitive with frontier models for long-context engineering tasks, not just code generation.