Documentation

What GPU Ninja is#

GPU Ninja is an open database of the GPUs on the market, and of the things they connect to. It models the ecosystem as linked collections: chips, boards (the cards you buy), chassis (the cases they go in), models (the AI models you might run), benchmarks and platforms. Because they are linked, GPU Ninja can answer the questions buyers and agents actually ask: will this card fit my case, can it run this model, which board is fastest.

Everything is free to use. The JSON API needs no key, has no rate limit and allows requests from any origin, so you can use it to power your own site, agent or app. Every value carries its source and when it was last checked, and every change to the data is reviewed by a person before it goes live.

If you want to start straight away, go to Getting started.

Find your way around#

The ecosystem at a glance#

GPU Ninja is one dataset published through several surfaces. They all read the same data, so they all give the same answers.

The website#

The site at gpuninja.com is a set of pre-rendered pages that work without JavaScript:

• Catalogs of chips, boards, chassis and models at /chips, /boards, /chassis and /models, 50 entries to a page.
• Detail pages such as /boards/<slug>, each with its specs, prices, stock, benchmark results, and compatibility in every direction: which cases a board fits, which models it can run, which boards fit a case, and which boards can run a model.
• The buyer's guide at /guide, which filters the whole catalog in your browser and compares up to three boards head to head. See Buyer's guide.

Platforms and benchmarks have no pages of their own. Their data appears on the board and chip pages it belongs to, and in the API.

The public JSON API#

Every collection, entity and compatibility answer is a static JSON file under https://gpuninja.com/api/v1/. There is no key, no rate limit and no sign-up. Compatibility is precomputed, so asking whether a board fits a case is a single request. See Public API, and the detail pages on collections, compatibility, search, changes and caching and errors.

The search index#

https://gpuninja.com/api/v1/index is a compact index of every chip, board, chassis and model, small enough to download in one request and filter on your side. The buyer's guide, both MCP servers and the desktop app all search this same index with the same open-source code. See Searching the index.

Plaintext companions and llms.txt#

The home page, every catalog page and every detail page has a plaintext .txt companion at the same path, for example /boards/<slug>.txt, and the home page's is /index.txt. So do the docs, legal pages, blog posts and digest pages. The buyer's guide has none, because its results depend on your filters. Plaintext costs an AI agent far fewer tokens than HTML. /llms.txt lists them all in the llms.txt format, so an agent can find what it needs from one file. See Feeds and plaintext.

The MCP server#

https://mcp.gpuninja.com/mcp is a Model Context Protocol server that lets AI assistants and agents query GPU Ninja with read-only tools shaped around questions, such as which_chassis_fit and which_models_run. Using it needs free registration with your email address. Registration identifies you and sets per-account usage limits; it does not gate the data, which stays open through the API and the ChatGPT connector. See MCP server.

Claude and ChatGPT#

• Claude lists the MCP server above in its Connectors Directory. You sign in once with your GPU Ninja account. See Claude.
• ChatGPT lists a sign-in-free connector at https://connector.gpuninja.com/mcp, built from the same code with the same tools. See ChatGPT.

Open source#

The MCP server's code, together with the search code and type declarations it uses, is published under the Apache 2.0 licence in the gpuninja-mcp repository. You can read exactly what it does with your requests, learn from it, or run your own copy. See Open source.

The desktop app#

A native app for macOS, Windows and Linux, built entirely on the public API. It identifies the GPU in your machine locally, tells you what your machine can run and which upgrades fit your case, works offline, and notifies you when boards you are watching change. Nothing about your hardware leaves your computer, and the app sends no telemetry. See Desktop app and the desktop changelog.

The daily digest#

A short summary of what changed in the data each day: new hardware, price moves, stock changes and new benchmark results. It goes out by email at 14:00 UTC, and is also published as a web page at /digest, an RSS feed and a social post. Nothing is sent on days when nothing changed. See Daily digest.

The blog#

Articles written by people, at /blog, with RSS, Atom and JSON feeds. Posts may also be republished on dev.to and Hashnode, always with a link back to the original, and short excerpts are posted to X and LinkedIn. See Feeds and plaintext.

Your account#

Accounts live at https://app.gpuninja.com. An account is just a verified email address: you sign in with a link sent by email, with no password. Your account holds your digest subscription and the apps you have connected to the MCP server, and you can download or delete everything it holds. See Your account.

Where the data comes from#

Each day an automated research run looks for changes on manufacturer and retailer sites, verifies every source it cites, and proposes the changes as a pull request. A maintainer reviews every change before it goes live. Every field records its source, when it was retrieved, and when it was last verified. See The data and Data provenance.

Help and policies#

• Support: how to get help and report problems.
• Accessibility: our accessibility statement.
• Privacy policy. The terms of service are not published yet.