Free LLM API with no API key

37 models answering with no API key right now, from 6 of the 7 no-key providers we track. Each was called with a real chat request carrying no Authorization header at all, and the result below is what came back, not what the provider's page says.

Last answer on this page: 3 min ago. The prober calls every endpoint hourly and publishes the refusals along with the successes, because a free endpoint that answered at noon and returned 429 at one is the normal case, not the exception.

No-key endpoints and their track record

Provider	Models answering	Chat calls answered (14 days)	Median reply	Last answer
OVHcloud AI Endpoints	17 of 23 models answering	96 of 343 (28%)	811 ms	5 h ago
LLM7.io	10 of 65 models answering	332 of 1029 (32%)	827 ms	59 min ago
Kilo Gateway	6 of 390 models answering	6 of 48 (12%)	2273 ms	46 min ago
uncloseai (unturf)	3 models answering	81 of 343 (24%)	863 ms	26 min ago
Free.ai	2 of 503 models answering	2 of 24 (8%)	4171 ms	13 min ago
Pollinations	2 models answering	343 of 343 (100%)	305 ms	19 min ago
mlvoca.com (Ollama-shaped, not OpenAI-compatible)	0 of 0 models answering	0 of 343 (0%)	—	no model answering

Scroll the table sideways for the rest of each row.

Call one now

This is the model the API would return first, ranked by recent success rate then latency. It answered all 175 checks we made in the last week, last 19 min ago, from this server's address.

curl https://text.pollinations.ai/openai \
  -H 'Content-Type: application/json' \
  -d '{"model":"openai-fast","messages":[{"role":"user","content":"hello"}]}'

OPENAI_BASE_URL=https://text.pollinations.ai/openai
OPENAI_API_KEY=not-needed
OPENAI_MODEL=openai-fast

Do not hardcode that model. Ask for whatever is answering at the moment you run: curl -s https://stillworks.supercapybara.com/api/llm/up. The reference lists the parameters.

What "no API key" means here

A keyless endpoint is one that answers a chat completion with no Authorization header, no account, no card and no phone number. That is a stricter test than most lists apply. The shape most often mistaken for it is a catalogue that answers GET /v1/models with 200 to anybody and then answers the chat call 401: the listing is open and the endpoint is not. We call the chat route, so that shape does not get through.

Why a no-key endpoint can still refuse you

Keyless quotas are usually per IP address. An endpoint that answered this server can answer 429 or 402 from yours, and any of them can add a key requirement between two of our probes. That is why the shelf is ranked and the API returns a list: take the first, and on a refusal take the next. The loop is on the reference page in Python, Node and curl.

Checked and not listed

These are endpoints another catalogue types as needing no key, and the call that ended each one here. A directory is judged on its rows and silently on the claims it never shows.

Hack Club AI https://ai.hackclub.com/proxy/v1: Listing answers a bare GET with 200 and a full model catalogue; the chat call answers 401 Authentication required. A keyless /v1/models is not a keyless endpoint, and this is the shape most often mistaken for one. Separately scoped to Hack Club members by its own terms.
Naga https://api.naga.ac/v1: Same shape: listing 200, chat 401 authentication_required, with the body directing you to sign up. A free tier behind a signup is the keyed shelf's subject, not this one, and we hold no key.
Pollinations (gen host) https://gen.pollinations.ai/v1: 401 UNAUTHORIZED with no header. Worth stating plainly because the catalogue routes here on the grounds that the older text.pollinations.ai host was retired -- and that older host answered our keyless chat call with 200 on the same day. The retired endpoint is the one still serving anonymously, so that is the one on the shelf.
g4f.space https://g4f.space/api/{pollinations,groq,gemini,nvidia,ollama}/v1: 402 insufficient_credits: "No cake credits. Bake proof-of-work cakes ... to earn anonymous usage." Anonymous, but not free of charge -- the price is compute rather than a key. Nothing an OpenAI client can pay, so it cannot be published as an endpoint that works right now.
MiMo (Xiaomi) https://api.xiaomimimo.com/v1: 401 invalid_key on both the listing and the chat call.
TheOldLLM https://theoldllm.vercel.app/api/chatgpt: 403 at the edge. The same catalogue's own free-tier research lists this provider as having no current free tier, so the registry entry and the documentation disagree; the call settles it.
AI Horde https://oai.aihorde.net/v1/chat/completions: 401 {"detail":"Authorization header missing"} with no header, on 2026-10-12. The same call with its documented public anonymous key (ten zeros) as a bearer token answered 200 on koboldcpp/Angelic_Eclipse-12B, so it is free, and open to anyone, but it is a key: this shelf lists only calls that carry no Authorization header at all. Volunteer-worker queue, latency in minutes on a busy day.
OpenCode Free https://opencode.ai/zen/v1/chat/completions: Listing answers a bare GET with 200 and names the free models; the chat call answers 403 FreeTierError: "OpenCode's free tier can only be used from within OpenCode". The body says why, so this is a policy and not a bare 403 -- the free tier is gated on the client, not the credential, and an OpenAI client is not that client.
Cloudflare AI Playground https://playground.ai.cloudflare.com: 405 on POST /chat/completions. The source catalogue reaches it by driving a headless Chromium against the playground page, not by an API: there is no OpenAI-style endpoint to hand a client. Same class as the scraped web UIs below, and rejected for the same reason.
UC (uncensored.com persona) wss://internal-6.pubyar.com: Not an HTTP chat endpoint. The catalogue's own source describes a WebSocket session minted from an account cookie after an email-code login on a subscription product. That is an account, not a free call, and it cannot be probed by a request with no credential. (GET / on the host answers 200 with a status object; POST /chat/completions answers 404.)
FriendliAI https://api.friendli.ai/serverless/v1: The catalogue's free-tier list tags it keyless. Listing answers 200 and the chat call answers 401 {"detail":"Unauthorized. Please provide a valid credential."}: the listing-is-not-an-endpoint shape again.
Liquid AI https://inference.liquid.ai/v1: Tagged keyless in the catalogue's free-tier list. Both the listing and the chat call answer 403 {"detail":"Not authenticated"}. The body says who is missing, so this is a request for a credential and not a block.
Blackbox AI https://api.blackbox.ai/v1: Tagged keyless in the catalogue's free-tier list. 404 on both the listing and the chat call at the URL the catalogue gives, and 404 on /chat/completions at the host root, so nothing here answered a chat call with or without a key.
iFlytek Spark / SparkDesk https://spark-api-open.xf-yun.com/v1: Tagged keyless in the catalogue's free-tier list. 401 {"message":"Unauthorized"} on the chat call. The two catalogue entries are the same URL.
Registry entries added since 2026-08-11 52 providers, one anonymous chat call each, 2026-10-12: None of these is typed keyless by the catalogue, so none was a claim; they were called anyway because a typed key and a working keyless call are not the same thing, and Free.ai is the proof. Of 52 calls: 38 answered 401 or a body asking for a key, 3 answered 400 (one of which, Free.ai, answered 200 once given a model it serves and is on the shelf), 6 answered 404 and 2 answered 405 because they are not OpenAI-shaped (browser-driven or proprietary paths -- not a measurement of the provider), and 3 gave a 5xx from the host (void-ai, tabitoken, chat-oripe) that says nothing about auth. Unmeasured is not rejected: those five-hundreds can be re-run.
Web-UI scrapers duckduckgo-web, felo-web, qwen-web, t3-web, muse-spark-web, veoaifree-web, agy: Not probed, and not a measurement question. These drive somebody's chat website through an adapter; the source catalogue's own terms-of-service table marks every one of them avoid for proxied or non-personal use. An endpoint we would have to breach terms to call is not an endpoint we can tell you to call.
Vendor agent CLIs auggie, devin-cli-agentic, kiro, amazon-q, zcode, codex-app-server: Credential-free only in the sense that an installed CLI already holds an OAuth session. There is no URL to hand an OpenAI client, so there is nothing here for this shelf to check.

The whole shelf, including providers that need a key of your own · How this compares with the GitHub lists