Michael JohnsonOn October 5, 2026, startup Ghost announced that it had emerged from stealth and raised an $11 million seed round led by Andreessen Horowitz. The company’s first product is Core, a $3,499 computer that Ghost positions as a home device for running local AI models and agents continuously. TechCrunch reported on the funding round and product, while Ghost’s website lists the price and specifications.
Core is designed to work with personal context: the company describes connections to apps, files, and devices. The device’s value therefore depends both on its hardware and on what data the agent can access, how it uses network features, and what actions it can take on the owner’s behalf.
What’s known about the computer
According to Ghost’s specifications, Core is equipped with an NVIDIA RTX PRO 4000 Blackwell SFF graphics card, an AMD Ryzen 5 7600 processor, 64 GB of DDR5 RAM, and a 1 TB SSD. For the graphics card, NVIDIA lists 24 GB of GDDR7 memory and 432 GB/s of bandwidth.
These figures describe the components. They don’t show how the finished Core performs on specific tasks. The available materials include no independent device tests, so generation speed, agent performance, power consumption, and noise levels have not yet been assessed.
Ghost founder Zain Jawaid said the company planned to begin shipping the first batch in the final week of October. As of October 6, that was a stated future timeline; TechCrunch had not yet reported that shipments had begun.
Local processing and network features
Ghost says that conversational AI and memory run on the device, and that Core can connect to apps and files and retain context. Some features, however, use the internet and external services.
Ghost’s privacy policy, which took effect on October 4, 2026, describes local storage and processing of conversations, files, content from connected apps, and memories. For online features such as web searches or calls, a request may travel from Core through Ghost’s gateway to a third-party provider. The company says its gateway forwards the request and response content without storing it; the external provider processes the request under its own rules.
For remote connections to Core, Ghost offers a relay that passes traffic between the device and the app. The policy also describes the option to replace the relay with a user’s own connection method. Local model execution and remote access are therefore separate parts of the system, and the data route depends on the integrations and features selected.
Screen history is disabled by default. If a user enables it, Core saves screenshots and recognized text. According to the policy, these records are deleted seven days after the last observation or sooner if the memory limit is reached. This expands the context available to the agent while also increasing the amount of information stored on the device.
Access to data and actions
Core is designed as an agent that accesses connected sources and carries out tasks. The broader its access to email, calendars, documents, and accounts, the more important clear permissions and control over the agent’s actions become.
In an interview with TechCrunch, Jawaid described a network filter under development that is intended to monitor the agent’s outgoing requests and block some actions. The published materials contain no independent technical assessment of this mechanism or confirmation that it is already included in the shipping version of Core. Ghost says it is willing to arrange an independent review of its managed gateway and relay; the sources reviewed contain no published results of such an audit.
Local storage determines where some information is processed. A broader assessment of security also depends on connected accounts, permissions, remote access, and the agent’s behavior. If the device is authorized to act through a user’s accounts, clear settings and boundaries on those permissions become part of the product’s value.
What the first shipments will show
Ghost sells Core as a ready-to-use system with preinstalled models and a software layer for running agents. The announced price and configuration are already known, but the available materials do not yet show how reliably the device completes tasks, how it behaves during network operations, or how clearly users can manage memory and permissions.
Independent tests of the first devices will make it possible to assess these aspects in practice. For now, the confirmed details are the funding announcement, the price, and the published specifications; Core’s practical value depends on how its software system performs and how well users can control agents’ access to data and their actions.