Agentic Engineering
Ghost Core's $3,499 local-agent test
Ghost Core puts personal agents on a $3,499 computer. Its 24GB GPU makes workload fit, context quality and control the purchase tests.
Treat Ghost Core as a personal-agent pilot with a hardware acceptance test: Ghost advertises the computer for $3,499, with models running on the device. Its 24GB GPU is a concrete constraint; buying the box is only the beginning of proving that the assistant earns its place beside an existing computer.
The purchase buys a bounded machine
Timing makes this a purchase decision rather than a distant concept. TechCrunch reports preorders opening Monday and shipping scheduled for October’s final week. That is a reported schedule. Buyers should obtain the delivery and acceptance terms attached to their order before making another workflow depend on the device.
The hardware boundary is unusually legible. Ghost lists an NVIDIA RTX PRO 4000 Blackwell SFF GPU, 64GB of DDR5 system memory, and 1TB of NVMe storage. NVIDIA independently specifies 24GB of GDDR7 GPU memory for that card. System memory and dedicated graphics memory are separate entries, not interchangeable labels for the same capacity. A procurement sheet should preserve that distinction instead of promoting the larger number into a claim about GPU-resident models.
Here is a reproducible budget figure: combine Ghost’s $3,499 system price with NVIDIA’s 24GB specification. $3,499 ÷ 24 = $145.79, or $146 of whole-system purchase price per GPU-memory gigabyte, rounded. This is neither the component price of memory nor a measure of inference efficiency. It attributes the entire purchase to one constrained resource so the buyer can ask whether the workload justifies paying for this particular package. It cannot rank unlike architectures or establish a cloud break-even point.
The important missing denominator is useful work. A price per completed, accepted task would require measured throughput, task quality and operating costs. None of those measurements comes from dividing the sticker price by memory. A comparison that quietly assumes every local response replaces a paid cloud response would also overstate savings. Some work may remain remote; some may be new activity that the owner would never otherwise have purchased.
Ghost’s product page makes a broader software pitch: continuous context, proactive assistance, no subscription, and local processing. Those are vendor claims to test, not independent findings about privacy or productivity. The useful pilot asks whether the assistant remembers the right information, forgets what it should, and chooses appropriate moments to act. An assistant that generates more interruptions could satisfy a technical definition of proactive behavior while making the owner less effective.
Our earlier PAIR analysis separated request routing from pooled model memory. Core calls for the same care with boundaries: possessing a dedicated computer does not by itself establish how a particular model, context setting or workflow behaves. Ask for results from the configuration that will arrive, then reproduce the important ones with the actual applications and permissions intended for daily use.
Buy the workflow only after it passes
Start with a narrow job whose outcome can be checked. A personal-document retrieval pilot, for example, could use a deliberately limited collection with known answers and explicit exclusions. The evaluation should distinguish retrieving the correct record from producing a plausible response. Include outdated information, contradictory records and questions the assistant cannot answer. These are proposed acceptance conditions, not claims that Ghost has passed or failed them.
Context deserves a deletion test because Ghost’s privacy policy says removing a source does not necessarily remove a memory derived from it. It also distinguishes copies on paired devices and backups from the information held on Core. Deliberately change a fact in the permitted collection, then check correction and removal through the product’s normal controls. This is a documented data-lifecycle boundary to evaluate, not a hypothetical accusation that the product secretly keeps everything.
Local inference also needs a precise meaning. The same policy says conversational AI and memory have no remote-model fallback, while online features can send task information to external services. Ghost says its gateway does not retain forwarded content; third-party retention follows the provider’s own terms. Map the services intended for the pilot before connecting sensitive material. Begin with retrieval or suggestions, specify which changes require confirmation, and check how access can be revoked.
The economics need a ledger rather than a subscription comparison. Record the purchase, setup time, any continuing services, maintenance effort and the value of accepted work. Measure power if electricity materially affects the proposed usage. Do not substitute the graphics card’s rating for the computer’s observed consumption. Likewise, do not treat time spent supervising the pilot as free simply because there is no separate software invoice.
The strongest case for Core is convenience. A buyer may willingly pay for an integrated assistant if assembling and maintaining an equivalent experience would consume scarce attention. That is a legitimate product benefit even if the hardware is not the cheapest way to obtain GPU capacity. The right comparison is the buyer’s current workflow plus the work required to change it, with a clear definition of what becomes easier.
The strongest countercase is underuse. A dedicated device has to justify being maintained after the novelty fades. Keep an observation log that includes abandoned tasks, unwanted interventions and manual corrections alongside successful results. A product that helps with a few recurring jobs may still be worthwhile, but the decision should rest on those jobs rather than a promise to understand everything about its owner.
Today’s lead asks advertisers to prove what their measurement setup actually observes. Personal agents deserve the same discipline. Consider a bounded Core trial if local processing and reduced setup effort solve a specific problem. Expand only after delivery, repeatable task success and understandable controls. Independent workload results and clear recovery behavior would strengthen the purchase case; unexplained failures or continuing dependence on manual correction would weaken it. The check to write is $3,499. The evidence to demand is a workflow worth keeping.