LLM API Check

Browser-only LLM capability & performance tester. Register any OpenAI-compatible model (OpenAI, DeepSeek, Qwen/DashScope, Volcengine Ark/Doubao, MiniMax, Zhipu GLM, Moonshot Kimi, Gemini, xAI, Mistral, Groq, OpenRouter, Ollama, vLLM, LiteLLM, …) and measure it from your browser — no backend, no analytics, API keys never leave the page.

• Performance: streaming / non-streaming TTFT, time-to-first-content-token, decode & end-to-end tokens/s, p50/p95, prompt-cache cold vs warm, with randomised prompts so caches cannot inflate results — plus scenario scores (real-time voice, chat UI, agent loops, batch).
• Capabilities: reasoning by default, reasoning_effort, every known "disable/enable thinking" dialect (reasoning_effort: none|minimal, thinking:{type:disabled}, enable_thinking, chat_template_kwargs, OpenRouter reasoning, Ollama think, Gemini thinking_budget), tool calling (auto / required / named tool_choice, parallel, streaming, round-trip, strict), structured output (json_object, json_schema, strict, nested), vision (base64 + URL), prompt caching (automatic + cache_control), parameter compatibility.
• Message formats: multiple / mid-conversation system messages, consecutive user or assistant turns, assistant prefill, tool calls without results, orphan tool results, content-parts arrays.
• Privacy: static site, strict CSP, no third-party requests; everything is stored in localStorage (keys optionally in sessionStorage) and can be wiped in one click. See /privacy.
• SEO: every route is prerendered to static HTML in English (/, the default) and Chinese (/zh), with hreflang (x-default → English), Open Graph, JSON-LD and a sitemap. A remembered language choice (header toggle or banner) redirects before first paint via the inline script in index.html; without one, Chinese-language browsers opening an English URL are sent to /zh the same way, while English-language browsers on /zh only get a one-time switch banner. Crawlers render as en-US with no storage, so they never see a redirect. Legacy /en/* URLs 301 to the root via _redirects.

Development

npm install
npm run dev                 # http://localhost:5173
npm run mock                # OpenAI-compatible mock server on http://localhost:8787/v1 (models: mock-fast, mock-thinker, mock-strict, mock-slow, mock-flaky)
npm test                    # unit tests (vitest)
npm run typecheck

To try the UI without real keys, add a Custom provider with base URL http://localhost:8787/v1 and any key (e.g. sk-test-1234567890), then add the mock-* models.

Build

VITE_SITE_URL=https://llmapicheck.dev npm run build

build runs the client build, an SSR build and scripts/prerender.mjs, which writes one HTML file per route and locale into dist/, plus 404.html, robots.txt, sitemap.xml, _redirects (legacy /en/* → /*) and _headers (CSP and other security headers).

VITE_SITE_URL (or Cloudflare's CF_PAGES_URL) is used for canonical URLs, hreflang alternates, Open Graph URLs and the sitemap. It defaults to https://llmapicheck.dev, the production domain; set it when deploying a copy elsewhere.

Deploy to Cloudflare

The site is 100 % static. Requests go from the visitor's browser straight to the LLM provider; nothing runs on Cloudflare except static asset serving.

Option A — Workers (static assets, recommended)

wrangler.jsonc is already configured (assets.directory = ./dist, not_found_handling = 404-page).

npx wrangler login
npm run deploy   # add the custom domain llmapicheck.dev in the Worker's Domains & Routes settings

Or connect the repository in the Cloudflare dashboard (Workers & Pages → Create → Workers → Import a repository) with:

• Build command: npm run build
• Deploy command: npx wrangler deploy
• Environment variable: VITE_SITE_URL = your public URL

Option B — Cloudflare Pages

npm run deploy:pages

Or via Git integration: framework preset None, build command npm run build, build output directory dist, environment variable VITE_SITE_URL. _headers and 404.html are picked up automatically.

Custom domain

Add the domain in the Worker/Pages settings and set VITE_SITE_URL to it before building so canonical URLs and the sitemap match.

Project layout

src/lib/llm        transport: OpenAI-compatible client, SSE parser, stream accumulator, error classification
src/lib/perf       performance suite: randomised prompts, runner, statistics, scenario scoring
src/lib/caps       capability & message-format probes (tests/*), registry, runner, reasoning dialects
src/lib/providers  provider presets (base URLs, CORS notes, reasoning dialects)
src/lib/store      zustand stores persisted to localStorage (keys in a separate store)
src/i18n           zh-CN / en catalogues (all UI copy and probe explanations)
src/pages, src/components, src/content   UI and long-form docs
scripts/           prerender, mock server, asset generation, Playwright e2e drive

Adding a provider preset

Edit src/lib/providers/presets.ts. A preset only needs a base URL; optionally add alternative regions, the auth header name (e.g. api-key for Azure), extra headers, suggested model IDs, the max-tokens parameter name and the reasoning dialects the provider documents.

Adding a probe

Create a CapTestDef in src/lib/caps/tests/*.ts, register it in src/lib/caps/registry.ts, and add name / desc / why for its id under caps.tests plus any new messages codes in both src/i18n/en.ts and src/i18n/zh.ts. The i18n unit test fails if a code or test id is missing a translation.

License

MIT