Kostenlose LLM- und Übersetzungs-API über Cloudflare Workers AI nutzbar

If you're building a side project and just want an LLM endpoint without wiring up a paid API key, Cloudflare Workers AI is worth a look. The free plan includes a daily allocation of "neurons" (Cloudflare's usage unit) for open models like Llama 3.1 8B and m2m100. That's plenty for prototypes, demos, and low-traffic tools. Limits are set by Cloudflare and can change, so check the current pricing page.

I packaged two small, tested Workers as an MIT-licensed free download: cf-ai-starter-lite.zip. The full code is below too, so you can copy it straight from this post.

name = "cf-chat-api" main = "worker.js" compatibility_date = "2026-10-01" [ ai ] binding = "AI" Enter fullscreen mode Exit fullscreen mode The [ai] binding is the only thing that makes Workers AI available as env.AI.

// Free LLM chat API on Cloudflare Workers AI // POST /chat {"message":"hi","history":[{"role":"user","content":"..."}],"system":"optional"} const MODEL = " @cf/meta/llama-3.1-8b-instruct "; const cors = { " Access-Control-Allow-Origin ": " * ", " Access-Control-Allow-Headers ": " Content-Type, Authorization ", " Access-Control-Allow-Methods ": " POST, OPTIONS " }; const json = ( d, s = 200 ) => new Response ( JSON. stringify ( d ), { status: s, headers: { " Content-Type ": " application/json ",... cors } }); export default { async fetch ( req, env ) { if ( req. method === " OPTIONS " ) return new Response ( null, { headers: cors }); if ( env. API_KEY && req. headers. get ( " Authorization " )!== `Bearer ${ env. API_KEY } ` ) return json ({ error: " unauthorized " }, 401 ); if ( req. method!== " POST " ) return json ({ error: " POST /chat " }, 405 ); let body; try { body = await req. json (); } catch { return json ({ error: " invalid JSON " }, 400 ); } if (! body. message || typeof body. message!== " string " ) return json ({ error: " message (string) required " }, 400 ); const history = Array. isArray ( body. history )? body. history. slice ( - 20 ): []; const messages = [{ role: " system ", content: body. system || " You are a helpful, concise assistant.

" },... history, { role: " user ", content: body. message }]; const out = await env. AI. run ( env. MODEL || MODEL, { messages, max_tokens: body. max_tokens || 512 }); return json ({ reply: out. response, model: env.

MODEL || MODEL }); }, }; Enter fullscreen mode Exit fullscreen mode Template 2: translation API 04-translate/wrangler.toml:

name = "cf-translate" main = "worker.js" compatibility_date = "2026-10-01" [ ai ] binding = "AI" Enter fullscreen mode Exit fullscreen mode 04-translate/worker.js:

// Translation endpoint using Meta m2m100 (100 languages) // POST {"text":"Hello","target_lang":"french","source_lang":"english"} or GET /?text=Hello&to=spanish const MODEL = " @cf/meta/m2m100-1.2b "; const json = ( d, s = 200 ) => new Response ( JSON. stringify ( d ), { status: s, headers: { " Content-Type ": " application/json ", " Access-Control-Allow-Origin ": " * " } }); export default { async fetch ( req, env ) { let text, to, from; if ( req. method === " GET " ) { const u = new URL ( req. url ); text = u. searchParams. get ( " text " ); to = u. searchParams. get ( " to " ); from = u. searchParams. get ( " from " ); } else { try { ({ text, target_lang: to, source_lang: from } = await req. json ()); } catch { return json ({ error: " invalid JSON " }, 400 ); } } if (! text ||! to ) return json ({ error: " text and target_lang required " }, 400 ); const o = await env. AI. run ( MODEL, { text, target_lang: to, source_lang: from || " english " }); return json ({ translated_text: o. translated_text, target_lang: to }); }, }; Enter fullscreen mode Exit fullscreen mode Deploy Put each pair of files in its own folder (or unzip the free download), then:

cd 01-chat-api npx wrangler login npx wrangler deploy Enter fullscreen mode Exit fullscreen mode Repeat in 04-translate.

curl -X POST https://<your-chat-worker>.workers.dev \ -H "Content-Type: application/json" \ -d '{"message":"Explain edge functions in one sentence"}' curl "https://<your-translate-worker>.workers.dev/?text=Hello%20world&to=french" Enter fullscreen mode Exit fullscreen mode Things I learned along the way Small models need tight system prompts. An 8B model drifts with vague instructions. Short, explicit rules ("answer in at most 3 sentences", "reply only with JSON") help a lot.

Put auth in early. A public workers.dev URL burns through your free allocation quickly if someone finds it. The chat Worker checks an optional API_KEY secret ( npx wrangler secret put API_KEY ), then expects Authorization: Bearer <key>.

Handle CORS, including the OPTIONS preflight, or browser calls will fail even though curl works. Test against the real API. The free download includes test/run-tests.mjs, which calls Workers AI over REST with your account ID and token.

Mocks wouldn't have caught model-output quirks. CF_ACCOUNT_ID =... CF_API_TOKEN =... node test /run-tests.mjs Enter fullscreen mode Exit fullscreen mode Full version (paid, optional) The free download is complete on its own.

I also sell a fuller pack ($7) that adds an OpenAI-compatible /v1/chat/completions proxy with streaming (point the OpenAI SDK or LangChain at it), a batch summarizer, a JSON extractor with schema retry, semantic search with BGE embeddings, a prompt library, and a setup guide: https://cf-ai-starter-shop.pages.dev

Not affiliated with Cloudflare. Questions welcome in the comments.