Your next PC,
engineered for you.
Precision hardware selection, constraint-aware optimisation and real compatibility analysis. Built for people who want to understand every part of their machine, not just the price.
- 8
- component categories
- 16
- compatibility checks
- 0
- LLM-selected parts
Every build is assembled, not suggested
Parts are chosen because they satisfy a hard compatibility rule and improve the score — the engine explains both for each one.
Studio reference imagery. Component part numbers and prices come from the reference catalogue.
The approach
Built around compatibility.
Optimised around you.
A parts list is easy. A machine that is coherent — thermally, electrically and mechanically — is not. SpecForge treats a PC as a constrained engineering problem: feasible first, then optimised for what you actually run.
Compatibility is a constraint, not a filter
Seventeen checks run on every partial candidate — socket, chipset, memory generation, clearance, power connectors, interfaces. A build that fails one is never shown to you.
The search is deterministic
Beam search over compatible partial builds. The same inputs against the same catalogue version always return the same machine, and the run reports how many states it explored.
Spend goes where it changes the result
Budget is allocated by workload demand, then performance stops improving at the point the workload is satisfied. Excess capacity is never presented as value.
How it works
Three stages, no black box.
Every decision is traceable to a rule or a score. Nothing is generated by a language model.
Define your workload
Budget, region, the applications you actually run and their relative weight. Resolution, quality target, frame-rate goal and any brand or clearance preferences.
Loads from a template or from plain language
Optimise your components
Beam search expands only compatible, affordable partial builds. Incompatible candidates are rejected on the rule they break; survivors are scored on demand fit, balance and headroom.
Deterministic — identical inputs, identical build
Understand your build
Eight components with the reason each was selected, a full compatibility report, estimated performance, power headroom, bottleneck analysis and upgrade paths.
Every figure traces back to a catalogue value
The engineering
The reasoning is part of the product.
A recommendation you cannot interrogate is a guess with a price tag. Each build ships with the checks it passed, the model behind each estimate and the search that produced it.
CPU socket
Verified from catalogue specifications
Memory generation
CPU and board agree with the selected kit
GPU clearance
Card length against case clearance
PSU connectors
Native 12V-2×6 or 8-pin adapter path
Storage interface
M.2 slot count and PCIe generation
…and twelve more checks, all listed in the build
Constraint-aware selection
Candidate parts are pruned by hard requirements before they can enter a build — never after.
Deterministic optimisation
Beam search with explicit width and pool limits, tie-broken by price then part number.
Compatibility analysis
Socket, chipset, memory, clearance, radiators, connectors and interfaces, each with a verdict and a reason.
Performance modelling
Benchmark-anchored estimates with resolution and preset scaling, plus 1% lows and the limiting component.
Upgrade simulation
Re-run the real optimiser against a raised budget or new preference and diff the result.
Report structure shown with representative check names. Actual verdicts, values and reasons are produced per build from the catalogue.
Starting points
Selected templates.
A template sets real builder inputs — budget, workloads, resolution and preferences. It never contains pre-selected parts; the optimiser still assembles the machine.
Transparency
Build with confidence.
SpecForge does not hide uncertainty behind a score. Where the catalogue cannot verify something, the build says so.
Sourced specifications
Every component record carries the specification values the checks are validated against, with a verification status and a source.
Checks you can read
Each of the compatibility rules reports pass, warning, fail or unknown. Unknown is never presented as compatible.
Reference pricing, stated as such
Prices are indicative reference values for the catalogue version, not live retailer listings — the build says so each time.
Explainable selection
For every part: why it was chosen, what it costs, what the alternatives were, and what changing it would do.
Catalogue
A versioned reference set, not a live scraper. Builds record the version they were generated against.
Estimates
Performance figures are modelled from benchmark anchors and labelled as estimates wherever they appear.
Make every component count.
No account needed to generate a build. Sign in when you want to save, compare and share it.