Perplexity CEO Aravind Srinivas announced the launch of the Personal Computer cloud AI agent service in a blog post on the X platform yesterday (March 11). Leveraging the continuously running Apple Mac mini, the service seamlessly integrates users’ local files, applications, and cloud AI to provide personalized intelligent assistance 24/7.
In terms of its operational mechanism, this service is similar to the “crayfish” OpenClaw solution, primarily relying on a continuously running Apple Mac mini. It deeply integrates the user’s local applications with the cloud-based Perplexity AI, creating a collaborative “local + cloud” working model. Check demonstration video below:
This AI agent is available 24/7, ready to handle complex tasks and take over workflows for users at any time.
Unlike traditional “command-execution” AI, the core logic of Personal Computer is to act as an “AI project manager.” Users do not need to issue cumbersome specific instructions to multiple AIs; they only need to tell them the final desired outcome.
Upon receiving the target, the system automatically breaks down complex tasks into multiple sub-tasks and assigns them to different “sub-agents” for execution.
In extreme cases, this “AI project manager” can even write software code to complete specific tasks and ultimately present the summarized results directly to the user.
Take image processing as an example: If a user has a folder of photos on their desktop and wants to publish them on a website, AI can automatically analyze the image content, rename the files according to specific rules, and crop them to a size suitable for web page display without user intervention. Furthermore, users don’t need to directly operate their Mac mini; they can remotely control this local AI system via the internet.
Regarding data security, Srinivas emphasizes the system’s security boundaries. While the Personal Computer client continues to run locally on the Mac, the heavy AI computations are still performed on Perplexity’s “secure servers.”
To prevent AI from abusing its privileges, the system mandates that all instructions involving sensitive operations must be authorized by the user twice. Furthermore, all AI operation records are stored by the system and equipped with a “kill switch” to handle emergencies.
