furybot

Steps

1. Pick a model by the job, not by size. General, code, maths and science, vision. Each one says what it is good at and — more usefully — what it is weak at, so you can choose without knowing what a quantisation level is.
2. Say what you want back. Naming the output shape — a list, a diff, a paragraph, a table, a one-line answer — does more for the reply than any amount of instruction about tone.
3. Switch models mid-conversation when the job changes. The thread survives the switch. Draft the argument on the general model, hand the code to the code model, keep going.
4. Ignore the reasoning block unless you need it. Models that think out loud produce paragraphs of deliberation before reaching the point. That is kept in its own collapsible block so the answer reads as an answer — open it when you want to see how it got there.
5. Turn on web search only for the questions that need it. Everything else stays on our hardware. Search cannot: answering what happened this week means asking the public internet. It is a per-message choice with a notice at the moment you make it, and the reply cites what it found.
6. Check anything that matters. Language models invent details — names, dates, citations, figures — and state them exactly as they state correct ones. Nothing is checked. Treat legal, medical and financial answers as the start of a conversation with a professional.

Tips

• Your account page estimates how many more replies are in this week's budget, based on your own recent conversations rather than an average nobody matches.
• The API is OpenAI-compatible, so an editor plugin, a script or a terminal client works against the same models with a settings change and no plugin from us. See the API docs.
• Transcripts are encrypted where they are stored and are yours to delete. Deleting a conversation is a real delete, not an archive.

Limits worth knowing before you start

• The chat models are third-party open-weight models. We add no political layer of our own and the refusal logic around them is ours and symmetric, but we cannot remove a lean baked into a model's weights — the current models are known to push back on criticism of the Chinese government, and we would rather name that than let you assume a neutrality we cannot deliver.
• It will be confidently wrong. That is not a bug we are close to fixing; it is what the technology does.
• Nothing you write is used to train anything, and nothing is sent to a third-party AI service — except a web search you explicitly asked for, which by definition reaches the public internet.
• Nothing is archival. Uploads are deleted as soon as the job finishes, generated files are swept automatically after about a day, and the prompt text is blanked on the same schedule. Download what you want to keep — this is a workshop, not an archive.
• Adult content is absent from this product rather than switched off. There is no setting, plan, administrator action or support request that introduces it, on the site or on the API.
• Four things are refused on every plan, each one traceable to a law rather than to our taste. They are listed in full, with the statute behind each, on the safety page.

Questions

Is chat included on the free tier?

Yes. Free and paid accounts both include it, drawing on the same compute allowance as the picture tools — so a quiet week of images means more conversation, and the reverse.

Is my conversation used for training?

No. It is not sent to any third-party AI service and is not used to train anything. Transcripts are encrypted where they are stored and are yours to delete.

Can I use it from my editor?

Yes. The API is OpenAI-compatible, so anything that speaks that protocol works against the same models.

Account, billing and retention questions are on the FAQ.