Thursday AI For All

A free, captcha-gated proxy for AI inference. No account, no API key to buy.

Why?

When I was a youngin coming up in the brand new world of AI, I was incredibly interested. But I had no money to buy inference tokens. Inference is now cheap enough, and I am now fortunate enough, that I can give away a small amount of AI tokens for others like me who are looking to learn and experiment. Please do not abuse this service for running agentic workflows -- use it for running small experiments. You have much better free coding agents at your disposal, like OpenCode and Command Code.

There was also a practical reason for building it - I wanted to have a way for people to try free AI services backed by open models that I was building as experiments for my current project, Thursday.

This service is backed by an OpenRouter API key with a few dollars in credit. If it runs out, I will add more. This is as much an experiment in cheap/open models as it is an experiment in human kindness.

API

Errors return { ok: false, error: { message, type, code } }. Successful inference responses are standard OpenAI chat completions.

Base URL: https://ai-for-all.trythursday.org/api/v1

`POST /issue/challenge`

Ask for a proof-of-work challenge. You get payload.challenge (c, s, d) and a payload.token. Solve it, then send the nonces back with the token. Limited to one request per minute per IP.

`POST /issue`

Redeem a solved challenge for an API key.

{ "token": "...", "solutions": [1, 2, 3] }

Returns payload.apiKey (starts with sk-th-, valid for 24 hours) and the expiry in the X-Key-Expires-At and X-Key-TTL headers. One key per 10 minutes per IP. A solved challenge can only be redeemed once.

Keys are meant to be disposable. It is fine to lose one: get another. The budget and the hourly limit below are what protect the service, not the key.

`POST /chat/completions`

OpenAI-compatible chat completions. Send messages; tools, streaming and the rest of the OpenAI body pass straight through. 25 requests per hour per IP, and per key. Limits: 16k input tokens, 16k output tokens, 32k context. max_tokens is clamped to whatever room is left.

curl https://ai-for-all.trythursday.org/api/v1/chat/completions \
  -H "Authorization: Bearer sk-th-..." \
  -H "Content-Type: application/json" \
  -d '{"model": "local", "messages": [{"role": "user", "content": "hi"}]}'

`GET /models`

Lists the model aliases and what they point at.

local · qwen/qwen3.8-flash (Qwen3.8 Flash), provider-default quant · cheap, the default
frontier · deepseek/deepseek-v4.1-flash (DeepSeek V4.1 Flash), fp8 only · bigger, for when local isn't enough

Help me

Service budget

The above is the global balance for all users. When the daily budget is reached, the service 5xxs till midnight.

If you think this is an interesting idea, you can kick a few bucks my way to help keep the service running. If you use AI For All and find it helpful or use it to get into a situation where you can keep the cycle going, please do! You can also host a copy yourself and keep all the good will for your own.

If you can't contribute financially or by forwarding the service - sharing helps! Show it to your friends. Build cool stuff. Leave me feedback on Thursday. Or send a nice email showing me what you built, or just say hi!