models.rip for agents and developers

Every machine-readable surface of models.rip: the graveyard.json dataset endpoint, markdown content negotiation, llms.txt, and the MCP server.

Everything on models.rip is public, unauthenticated and free to use. There is no API key, no sign-up and no rate limit worth the name. The dataset is CC BY 4.0 — use it, credit it.

The dataset endpoint

/graveyard.json is the entire cemetery in one document: every model grave and every benchmark grave, with the fields the pages are rendered from. It is a static file, cached for an hour, and it is the right thing to fetch if you want all of it. There is no paginated or per-record REST API, because at this size one file is a better interface than one.

Model graves carry slug, name, provider, modelId, aliases, born, died, deprecatedAnnounced, retirementScheduled, causeOfDeath, successor, benchmarks, contextWindow, pricing, epitaph, lastWords and sources.
Benchmark graves carry slug, name, born, died, causeOfDeath, frontierAtLaunch, frontierAtDeath, ceiling, ceilingBasis, successor, epitaph and sources.
died: null means a reserved plot: deprecation announced, model still answering. retirementScheduled is the shutdown date the provider published, when it published one.
Every grave cites at least one source, and every source is the provider’s own page.

The HTTP API

Three read-only JSON endpoints, described in OpenAPI 3.1 at /openapi.json. Every operation carries a unique operationId, a description, typed parameters and a response schema, which is what a model needs to build a call to one without being told how.

/api/graves — search. q matches the name, the model id, any retired snapshot id, the provider and the epitaph text. Filter with provider, cause and status; page with limit and offset. total is the size of the answer, not of the page.
/api/graves/{slug} — one grave in full. Takes a retired snapshot id too: /api/graves/gpt-4-0613 answers with the GPT-4 grave and says what it resolved from.
/api/plots — every provider plot, with buried and reserved counted apart.

Failures come back as JSON in one shape, never as an HTML error page: an error object with a code from a closed set, a message, a hint naming the next thing to try, and a link back to this page. A parameter outside its range is refused rather than quietly clamped.

Markdown instead of HTML

Every page on this site serves markdown to a client that asks for it. Send Accept: text/markdown to any page URL and you get the content without the 3D scene, the navigation or the markup around it; the same URL still serves HTML to a browser. Responses carry Vary: Accept, so a cache cannot hand one audience the other’s copy.

Every page also has a .md sibling for clients that would rather follow a link than set a header: /gpt-4/epitaph.md is the epitaph, and the homepage is /index.md. Both are advertised from the HTML with <link rel="alternate" type="text/markdown">.

The command line

There is a CLI over the same API, with no dependencies and no build step: npx models-rip search claude haiku, npx models-rip show gpt-4-0613, npx models-rip plots. --json prints the raw payload for piping. show takes a retired snapshot id as readily as a slug.

llms.txt

/llms.txt is the llmstxt.org index: what this site is, when an agent should reach for it, and where the machine-readable versions live. /llms-full.txt is every epitaph in the cemetery in one file, for when fetching 59 URLs is the wrong shape.

The MCP server

models.rip runs a Model Context Protocol server over Streamable HTTP at /mcp. Point any MCP client at https://models.rip/mcp — no authentication, no session to establish. Its server card is published at /.well-known/mcp.

search_graves — find graves by name, model id, provider or cause of death.
get_grave — one grave in full, with its epitaph, dates and sources.
list_plots — every provider with a plot here, and how many each has buried.
cemetery_overview — what the archive is and what is in it, in one call.

Stable addresses

Slugs do not change. A retired snapshot id redirects to the grave that buries it — /gpt-4-0613 is a 308 to /gpt-4 — so an id pinned in production three years ago still resolves. /sitemap.xml lists every page, and /robots.txt allows every crawler, training crawlers included.

Using it

Attribution is the whole licence: name models.rip and link back to it. If you are building something that leans on the dataset, say hello — not for permission, but because it is useful to know what the data is being used for.