A private language-model service. You sign in with Google, ask questions in the browser, and the answers come from models running on hardware we control — with the option to search the web when a question needs current facts.
Go to thinkapenny.ai and press Sign in with Google. There is no separate password. Once you are signed in the front page shows what you can reach.
Access comes in three levels, and the one you get is decided by your email address:
| Level | Can use | Limits |
|---|---|---|
| guest | Chat | 10 messages an hour, 40 in total |
| member | Chat | None |
| admin | Chat and the console | None |
Anyone with a Google account can sign in and becomes a guest — you do not need to be set up in advance. An administrator adds you to the list to remove the limits. If you are a guest, the chat shows how many messages you have left.
The main way in. Ask a question, get an answer, keep the conversation going. It can search the web when you turn that on, and you can give it standing instructions about how you want answers formatted. See the chat guide.
For administrators: which models are loaded, what a request actually did, who has access, and what any rented hardware is costing. See the console guide.
Everything the chat does is available programmatically at api.thinkapenny.ai/v1, which speaks the OpenAI chat-completions format. That means most existing tools and libraries work against it unchanged — point them at the base URL and give them a token. Ask an administrator for one.
curl https://api.thinkapenny.ai/v1/chat/completions \
-H "Authorization: Bearer YOUR_TOKEN" \
-H "Content-Type: application/json" \
-d '{"model":"qwen2.5:14b","messages":[{"role":"user","content":"hello"}]}'Questions are answered by one of several machines, and the service picks one for you. You do not have to care which — but it explains two things you will notice.
Some models are instant, some take a few minutes. Models on our own always-on hardware answer immediately. Larger models run on rented GPUs that are started when first asked for and shut down when idle, so the first request after a quiet spell can take two or three minutes. Ask again shortly and it will be there.
Different models are good at different things. Ask for a simple answer and a small fast model is fine. Ask for a fifteen-row table with specific columns and a larger model is noticeably better at holding the shape you asked for.
Conversations run on hardware we control and are not sent to a third-party model provider. The exception is web search: when you turn it on, your search terms go to a search service, and pages it finds are fetched. Everything else stays here.
If you are a guest and want the limits removed, or you need an API token, ask an administrator to add you. They can do it from the console, and it takes effect the next time you load a page — you do not need to sign out and back in.