About custompc.ai

custompc.ai turns a sentence like "gaming, about $1,000" into a parts list that fits together. The list is produced by rules and a database, not by a language model: a model is consulted only when the rules cannot read a message, and it never picks a part.

Data source

The catalog is a snapshot of PCPartPicker's public listings taken on July 23, 2025: 66,778 parts across 25 categories, of which about a quarter carried a price at that time. Prices on the site are those reference prices until live store pricing is connected. Every price observation is stored with its date, so the price history for a part grows over time.

Method

Scores. Each part gets a 0-100 score. Graphics cards use published relative performance at 1080p, 1440p and 4K with the GeForce RTX 5090 at 100. CPUs combine boost clock and an architecture factor for single-thread speed, with a diminishing-returns core count term for multi-thread. Memory scores capacity first, then speed and latency. Storage scores interface first, then capacity.

Budget split. Each use (gaming, video editing, databases, machine learning, and so on) has a weight per component and hard minimums such as 16 GB of memory or 12 GB of video memory. A primary use carries 70% of the weights and secondary uses share the rest; minimums take the strictest value. The budget is divided by those weights, the best-scoring compatible part inside each share is chosen, and leftover money goes to whichever upgrade buys the most score per dollar. A soft budget may go up to 15% over for a clear tier jump.

Compatibility. Socket, memory type, slot count and capacity, board form factor against case class, card length against case class, power supply wattage with 30% headroom, radiator size, and NVMe availability. Where the catalog lacks a field, a conservative class estimate is used and the result is marked as estimated.

Bottleneck. Each use has a pipeline of stages with expected score ratios. A stage whose score falls below its expected share is loaded above 100%; the most overloaded stage past 115% is the bottleneck, and the tool looks up the cheapest compatible part that would relieve it.

Known gaps

The snapshot lists CPU sockets by architecture rather than per part, omits board memory type on some Intel boards, and has no cooler socket lists, case clearances or power-supply connector counts. Those are estimated from names and classes. The project's to-do list tracks sourcing them properly.

Affiliate disclosure

Store links may carry an affiliate code. The recommendation engine has no access to that information and cannot favor a store or a part because of it.

For AI agents

A remote MCP server is at https://custompc.ai/mcp and the same tools are exposed in-page through WebMCP where the browser supports it. See llms.txt. Agents are asked to present parts with the returned custompc.ai links.