A second opinion before you buy.
PC builds fail in expensive, predictable ways — a 360 mm radiator that the case can't actually mount, a PSU whose transient-spike rating trips under a modern GPU's power excursion, RAM that the board only runs at JEDEC speeds. Greenfield catches those before the order goes through.
We're not a retailer. We don't take affiliate cuts. We don't store your build. The URL is your build — share it, lose it, come back to it; that's the entire persistence layer.
Three flows. One source of truth.
Substitution check
A shopkeeper offered you a different SKU than the one you asked for. Paste both. We diff every spec that matters and tell you whether you're being up-traded, down-traded, or quietly handed an inferior part at the same price.
Run a check →02Build audit
Paste the full quoted build. Our compatibility engine runs every part against every other part across four layers — physical, electrical, performance, thermal — and surfaces what an experienced buyer would push back on.
Audit my build →03Budget recommendation
Set a budget and a use case. We return a tier-balanced parts list so no single component starves the rest. The budget is yours; the bottleneck math is ours.
Plan by budget →The model writes the prose. It doesn't decide anything.
Most chatbots will hedge a build is “mostly compatible” when there are four critical issues, or confidently say a 340 mm GPU clears a 330 mm case (it doesn't — 10 mm of overhang means the side panel won't close). We don't use the LLM as an oracle.
Catalog
Hand-curated specs scraped from manufacturer pages — only 2022+ parts, only what the vendor officially publishes. No retailer scraping, no benchmark inference.
Compatibility engine
Pure-function rules over typed spec fields. Every rule is operator-reviewed and labelled by severity (critical / power-real / performance-real / advisory). Cases are physical, not vibes-based.
Buyer-protection kernels
Above the compat layer: deterministic checks for substitutions, downsells, K→F swaps, RAM speed quietly cut, PSU transient ratings. These are the patterns we’ve seen shopkeepers exploit.
Language model
Only after the deterministic layer has produced findings does an LLM see anything. Its job is to write the explanation in plain English. It can’t hallucinate spec numbers because it never has to invent them.
The opinionated parts.
- Sell your dataNo accounts, no tracking pixels, no localStorage. One cookie for the chat session ID, that's it.
- Take affiliate kickbacksRecommendations would warp the moment someone's paying us per click.
- Hallucinate specsIf our catalog doesn't have a number, we say so. A null is a known unknown; a wrong number is a confidently wrong rule firing.
- Pretend benchmarks are universalA "tier 88" CPU is excellent for one workload and overkill for another. We surface what the tier score does and doesn't mean on a separate page so it's not mistaken for a value verdict.
Because the cost of a bad build is paid in cash, not feedback.
PCPartPicker catches the obvious incompatibilities — wrong socket, wrong form factor. It can't catch the ones that hurt: a PSU whose transient rating trips under modern GPU power excursions, an air cooler that overhangs the first DIMM slot on the board you bought, a motherboard whose VRM throttles the high-end CPU you spent two months saving for. Those failures don't show up at checkout. They show up six months in, when your machine is rebooting under load and you can't tell why.
We built the tool we wanted before we built our last machine. It's free, it'll stay free, and the catalog plus the rule set will keep growing.
Try a flow.
Three minutes, no login. The URL captures everything; close the tab whenever.