Ghost Core: a $3,499 computer for AI agents, explained

Ghost Core is a $3,499 screenless desktop that runs AI agents and open-weight models on the device instead of in the cloud. It uses an Nvidia RTX Pro 4000 SFF Blackwell GPU, and Ghost raised an $11 million seed round led by Andreessen Horowitz to build it. The interesting part is not the GPU. It's the software around the agent: a firewall on outgoing requests and encryption keys that stay with the user.

#What Ghost announced

Ghost came out of stealth on October 5, 2026. SiliconANGLE reported an $11 million seed round led by Andreessen Horowitz, with Abstract, Audacious Ventures, Nova and SV Angel. The CEO is Zain Javaid, who is 19.

The product is called Core. Ghost describes it as a "brain in a box". It has no screen and no consumer operating system like Windows or Ubuntu. You reach it through a phone app and a web app.

The first batch sold out at $3,499 each. Ghost has not said when the next batch opens.

#What's inside

Ghost has named one part: an Nvidia RTX Pro 4000 SFF Blackwell GPU. It has not published the CPU, RAM or storage.

The GPU is a known quantity. Retail listings and Puget Systems give it 24 GB of ECC GDDR7 memory and about 432 GB/s of bandwidth. It draws 70 W and needs no extra power cable. It's a small workstation card built for quiet, compact machines.

Retail listings put the card alone between about $2,000 and $3,500. So most of Core's price is the GPU. That makes the $3,499 box a fair deal on parts, before you count the software.

24 GB is enough for one mid-sized model at a time. A 30B-class model fits at 4-bit quantization with room for context. A 70B model does not fit on the card.

#The models

SiliconANGLE lists Qwen 3.8, a smaller Qwen 3.8 variant and Gemma 4 31B as preloaded. Users can download more from Hugging Face. RuntimeWire's coverage names Gemma 4 31B too, so treat the exact list as unconfirmed until Ghost publishes it.

These are good open-weight models. They are not frontier models. For long coding sessions with many tool calls, a 30B-class model will do less per step than a frontier model behind an API.

#What "fully local" means here

Ghost's pitch is privacy. Javaid told SiliconANGLE that "all of the source code, logic, and model weights live on the device." The user holds the encryption keys, and Ghost says nobody else, including Ghost, can read what's stored.

Two details matter more than the slogan:

• A firewall watches the agent's outgoing requests. That's the right control. An agent that reads untrusted content can be told to send your data somewhere. Blocking outbound traffic by default is how you stop that.
• Useful actions still reach out. RuntimeWire notes that "useful actions still depend on connected accounts and some outside services." Core also takes over-the-air updates. So "local" means the model and your data stay on the box. It doesn't mean nothing leaves.

That's a sensible design. It's also the same set of controls you'd want for any agent, on any machine.

#What it means for coding agents

Core is a consumer product. It's for a personal assistant that manages your accounts, not for a coding agent working in a repository. But the problems it solves are the same ones coding agents have:

1. Where the model runs. Local models keep code and prompts on your machine. Frontier models behind an API do more per step.
2. What the agent can reach. An agent with shell access and an open network can leak secrets or source code.
3. Where the keys live. If the agent can read your API key, a prompt injection can send it out.

A dedicated box answers the first question with hardware. The second and third are software questions. They need answers whether the agent runs on Core, a Mac mini or a rented server.

#Getting the same controls on hardware you own

If you already have a machine with a decent GPU or unified memory, you can get Core's main properties with open-source software.

Hotcell runs each agent task in a sandbox: a Docker container, a Firecracker microVM on Linux, or an Apple Virtualization microVM on macOS. Turn on its egress control and outbound traffic is denied by default, enforced by the kernel on Linux and in microVMs. Its gateway gives each sandbox a short-lived, spend-capped, revocable token instead of your real API key, so a compromised agent can't take the key with it.

Autopilot is the coding agent harness on top. It works with local models through Ollama or LM Studio, or with 25+ providers using your own key. You can mix them: a local model for cheap steps, a frontier model when the task needs it.

dl sets both up with one command, on this computer or on your own server over SSH. See the install guide.

If you're choosing hardware, our mini PC comparison covers Mac mini, Strix Halo boxes and DGX Spark by memory, bandwidth and price.

#Who should buy Core

Core makes sense if you want a private assistant without setting anything up, and you're fine with what fits in 24 GB. The phone app and a managed software layer are the product.

It makes less sense if you're an engineer who wants a coding agent. You'd be paying for a consumer interface you won't use. You'd also be limited to models that fit on one 24 GB card, with no screen or general-purpose OS to work in.

Either way, Ghost's funding is a useful signal. Investors are now paying for the idea that agents should run on hardware the user controls. The firewall and the user-held keys are the parts worth copying.

#FAQs

#How much does Ghost Core cost?

$3,499. The first batch sold out, and Ghost has not said when it will take more orders. You can register interest on its site.

#What GPU does Ghost Core use?

An Nvidia RTX Pro 4000 SFF Blackwell, a 70 W workstation card with 24 GB of GDDR7 memory. Ghost has not published the CPU, RAM or storage.

#Does Ghost Core work without the internet?

The models and your data run on the device. But connected accounts and some outside services still need the internet, and the box takes over-the-air updates when online.

#Can Ghost Core run a coding agent?

Ghost hasn't positioned it for that. It has no screen or general OS, and it's managed through a phone and web app. For coding agents, a general-purpose machine running a harness like Autopilot is more practical.

#What model sizes fit on 24 GB of VRAM?

A 30B-class model at 4-bit quantization fits with room for context. A 70B model does not fit on a single 24 GB card.

#How do I sandbox an AI agent on my own computer?

Run it in an isolated environment with outbound traffic blocked by default, and keep real API keys outside it. Hotcell does this with containers or microVMs, optional default-deny egress and a gateway that issues short-lived tokens.