Three surfaces, all read-only, all over the same catalog: the JSON API for code, the llms.txt endpoints for LLMs and answer engines, and the command-line client for the terminal and agents.
Read API (/api/v1)
JSON over HTTP. No API key. Responses are cached for an hour and respect source visibility, so hidden sources never appear. Start at /api/v1, which lists every endpoint.
GET /api/v1- index of all endpointsGET /api/v1/search?q={query}- full-text search across evals, tools, models, papers, leaderboards, organizations, and peopleGET /api/v1/{evals|models|tools|leaderboards|organizations|people|capabilities|papers}/{slug}- one entity with its metadata, scores, and relationshipsGET /api/v1/papers/{slug}/textand/pdf- a paper's full text or PDF, where its license permits redistribution
curl "https://sophon.at/api/v1/search?q=swe-bench"
curl "https://sophon.at/api/v1/evals/swe-bench"
curl "https://sophon.at/api/v1/models/gpt-5"For LLMs and agents
Following the llmstxt.org convention, two plain-text endpoints give a model the catalog as context:
- /llms.txt - a curated, one-line-per-entity map of the catalog, linking to the API
- /llms-full.txt - the expanded companion, with descriptions, top scores, and relationships inlined so a model can answer without fetching each page
AI crawlers (GPTBot, ClaudeBot, PerplexityBot, and others) are allowed in robots.txt, and /api/v1 is open to them.
Command-line client (sophon)
A zero-dependency Node CLI (Node ≥ 18) over the same API. Hits production by default, so there is nothing to configure.
npm install -g sophon-at # then use `sophon …`
# or run once, no install:
npx sophon-at search "swe-bench"sophon search <query> [--type T] [--per N]sophon get <type> <slug>- any of evals, models, tools, leaderboards, organizations, people, capabilities, paperssophon paper <slug> --text | --pdfsophon api·sophon help·sophon version
Built for agents. Output is the human summary on a terminal and JSON when piped (or with --json), so scripts and agents get structured data with no flags. sophon help --json returns a machine-readable manifest of every command, flag, and example. Errors are {"error":"…"} in JSON mode; exit codes are 0 ok, 1 runtime/remote error, 2 usage error. SOPHON_BASE points it at another host (e.g. a local dev server).
sophon help --json # machine-readable command manifest
sophon get model gpt-5 | jq .scores
sophon search "swe-bench" --json | jq '.results.eval[].slug'Source and issues live in the repository; the package is sophon-at on npm.
Terms
- Read-only and unauthenticated; please keep request volume reasonable.
- Catalog content carries its upstream licenses - see attribution and the content policy.
- Hidden sources are excluded everywhere, including the API and llms.txt.
Back to About.