Startup Ghost unveiled Core, a computer designed to run AI agents capable of carrying out tasks on behalf of users, alongside its announcement that it had raised $11 million in a seed round led by Andreessen Horowitz. Abstract, Audacious Ventures, SV Angel, and Nova also participated in the round.
The company is led by 19-year-old Zain Javaid, as part of a founding team that includes 19-year-old Nicholas Chua, 24-year-old Yifei Chen, and 23-year-old Gautam Sharda. Ghost says Core represents what it describes as a “brain in a box,” but it does not include a screen; users interact with it through a phone or the app, with a voice mode available through the Core app.
A Computer Built for Agents, Not a General-Purpose Device
Core includes an Nvidia RTX Pro 4000 SFF Blackwell graphics processing unit, integrated into its $3,499 price, along with the company’s own software layer, browser, and file system. The device comes with three models preinstalled: Qwen-3.8-Next, Qwen-3.8-27B, and Gemma-4-31B, with the ability to install additional models from Hugging Face or load models owned by the user.
The difference Ghost proposes compared with a general-purpose computer such as a Mac Mini is that Core’s hardware and software were designed together to run agents, rather than leaving users to set up an AI environment themselves. The device can process information and tasks continuously and without repeated prompting, and it can also receive model updates over the internet or enable automatic updates.
Data and Memory on the Device
Ghost says that the models and personal memory run locally, and that personal information is encrypted so that no cloud provider or third party, including Ghost, can access it in a meaningful way. Core was also designed to keep working even if the company shuts down, since the source code, logic, and model weights are on the device rather than in external services.
The company is working on a firewall that monitors every network request originating from the computer. If it detects that an agent is attempting to carry out an unauthorized action or export data that should not be shared, it is supposed to reject the request and flag it for user review. Users can also set the level of autonomy granted to agents.
What Changes in Practice?
Ghost is not merely selling a more powerful device; it is attempting to turn agent operation from an intermittent software setup into an always-on device that keeps the user’s life context locally. The example Javaid gave was an agent that connects an elevated resting heart rate, staying up late to work, and seasonal context, then sends an alert about the possibility that the user is ill.
But this model places the burden of trust on the device itself: the more access it has to applications, user data, and credentials, the more critical control and review mechanisms and the isolation of network requests become. The agents’ practical performance, memory accuracy, firewall limitations, and actual user experience remain open questions that the announcement alone does not answer.
Price and Availability
The units are currently being built at Ghost’s facility in San Francisco. Preorders for the first batch begin October 5, and shipping is scheduled to start in the final week of October. The article did not disclose broader details about production quantities or the markets available outside the United States.