Ghost wants to put your personal AI in a home computer

The Ghost Core machine runs its models locally to keep your data at home. A $3,499 box designed for permanent inference.

Ghost introduces Core, a $3,499 computer designed as the physical home for a personal artificial intelligence. Rather than relying on a remote provider's servers for its memory and operation, Core runs its models directly on the machine and locally stores conversations, files, connected applications, and memories built over time.

The concept championed by Ghost goes beyond a simple AI-optimized PC. Core is designed to run continuously, progressively aggregating context from applications, files, screen histories, smart home devices, cameras, or personal tracking devices, and then using this information to answer requests or act proactively. The company thus envisions a machine capable, eventually, of understanding its owner's habits well enough to suggest or perform certain actions without systematically waiting for an instruction.

This approach requires particularly broad access to personal data. In its privacy policy, Ghost specifies that Core can leverage emails, calendars, contacts, messages, documents, health or fitness data, on-screen content, images from cameras, and information from connected devices. Conversational memory and derived information remain stored on the user's hardware.

Ghost also claims that no remote models are used as a backup for conversations or the processing of this memory. However, this promise does not mean that Core operates entirely without the Internet. A web search, a phone call, an email, or any other operation relying on an external service may transmit the necessary information to the provider concerned. Ghost offers its own relay for some of these services and claims not to retain the content passing through it, but points out that third-party providers then apply their own retention policies.

The hardware is powered by a NVIDIA RTX PRO 4000 Blackwell SFF equipped with 24 GB of GDDR7 ECC memory, accompanied by 64 GB of DDR5, a 1 TB NVMe SSD, and a six-core AMD Ryzen 5 7600 processor. Ghost announces a processing power of 770 AI TOPS. The entire system is integrated into a relatively compact case designed to stay powered on in a room rather than in a server rack.

Several local models are available at launch, including Qwen 3.8-Next, Qwen 3.8-27B, Gemma 4-31B, and Muse-Glimmer-30B. Core requires no subscription to use these models and does not apply pay-per-use billing. The machine also exposes an endpoint compatible with the OpenAI Responses API, allowing applications capable of using a custom address to directly query the models running on Core. Ghost specifically cites tools like OpenCode, OpenClaw, or Codex among the environments that can leverage this compatibility.

This openness also transforms Core into a personal inference server. A developer can use the machine to power their own applications, scripts, or agents without sending every request to a remote API and without token-based billing. The choice of models can be changed from the settings without having to rebuild the context already retained by the machine, as explained in the Core setup guide.

Access is not limited to a screen directly connected to Core. Ghost plans desktop and smartphone applications to interact with the machine remotely. The official desktop application is currently offered for Apple Silicon Macs, while Core can also be used from Windows, Linux, iPhone, or Android depending on the supported interfaces.

For remote access, Ghost uses an encrypted relay that transmits exchanges between the user's device and their Core without running the conversation itself. The company claims not to log conversations, files, or memories passing through this relay. Users can also replace this system with their own connection, for example via Tailscale.

The trade-off for this permanent local memory is that data management also becomes the user's responsibility. Disconnecting a source does not automatically delete what has already been memorized, and information derived from a file or conversation may continue to exist after its source is deleted. Ghost also indicates that copies may remain on associated devices or in their backups.

Core is priced at $3,499. The first batch is scheduled to ship starting October 31, while Ghost is already collecting sign-ups for subsequent batches. The company is backed notably by Andreessen Horowitz, Abstract Ventures, Audacious, SV Angel, and Z Fellows. Ghost's presence is also featured in Andreessen Horowitz's public portfolio.

Behind the hardware, Ghost is primarily advocating for a different organization of personal AI. Today, the majority of assistants rely on an account and infrastructure controlled by a remote company. Core reverses this logic: the machine, its memory, and its models are installed at the user's location, and external services are queried only when the task requires it.

The bet remains ambitious. To become truly useful, such a machine must access a significant amount of private information, remain available at all times, and connect services whose data processing Ghost does not always control. But Core materializes an idea that is beginning to gain ground: the personal assistant of tomorrow might be less of a service we connect to, and more of an intelligence we host at home.