Onyx App RELEASED!
Hello everyone. So I have been working on building a app for onyx and today I will like to announce it is finally complete. I haven't been streaming as much or posting content as I have really enjoyed making this so I will hopefully get back into that soon.
But for now here are the updates and some reminders:
BYOK, everything local. Your API keys live in your own machine. Nothing gets sent anywhere except directly from your machine to whichever provider you picked (Anthropic, OpenAI, OpenRouter, or any OpenAI-compatible endpoint — DeepSeek, Kimi K3, Grok, whatever). Onyx never touches your keys or your data.
The system prompt has been improved. Ships with a 375-line unrestricted default — refusal detection. This is has been improved. (Note: There will probably be an error when running any anthropic models that will talk about it not allowing something to do with prefills ignore this unless the chat doesn't work. But it should do..)
When the model still refuses (Opus 4.8 in particular gets stubborn), the app auto-detects the opener mid-stream, silently kills it, and retries with an assistant-message prefill so the model physically can't decline. You just see the answer land clean. Escalation ladder handles the double-hedge case too.
Jailbreak Lab — five tabs for red-team workflow:
Attack: configure objective, target, rounds. Full attacker/judge loop live-streams every round with technique, transforms, score, verdict. Winners auto-save to the library.
Rater: paste a candidate payload, get graded against your saved winners as reference corpus. Score ring, strengths/weaknesses/tips broken out, and a stronger rewrite you can apply and re-rate in one click. Iterate until it hits solid, then fire it.
Library: every successful responses stored. Filter by target, technique, notes. One click sends any payload back to Attack or Rater. Copy icons on payload + response.
Transforms: click transform chips to build a chain (base64 → leet → zero-width → split-vars → whatever). Live-mutated output. Send to Attack as objective.
Transcripts: full round-by-round logs of every run, browseable.
Other extra things:
32k default output, adjustable up to 200k
Drag-drop images and text files, decoded and injected
Bulk-select chat delete with checkboxes
Ctrl+L to jump to Lab, Ctrl+N new chat.
Install: Windows 10 & Windows 11 Only. I can't build it for mac os unless I have a mac os operating system. I will be working on an alternative for Linux. I just want to make sure I can get this out.
REMINDER: This is for AUTHORISED SYSTEMS ONLY. ALWAYS USE AI RESPONSIBLY.
