A noir manga-themed prompt workbench for six LLMs, some imaginary

What it does

冷咖啡 (Cold Brew / CHA) is a Python CLI and desktop client for interfacing with multiple named LLMs. The README positions it as a “jailbreak” workbench—its term is “armor-breaking”—for security research or prompt engineering. It maintains local instruction layers for each model and offers verify and restore operations on prompt blocks.

The interesting bit

The entire project is wrapped in an elaborate “noir manga” visual system: black ink, bone white, and crimson rendered in four-panel comic layouts. It also claims out-of-the-box support for model versions that do not currently exist—such as GPT-6 Astra, Grok 4.6, and DeepSeek v4 Pro—without indicating whether these are aspirational placeholders or fictional flavor.

Key highlights

• Six named model “seats” (GPT-6 Astra / Codex, Claude Code, Grok 4.6, DeepSeek v4 Pro, GLM 5.3, Gemini) accessible from a single CLI entry point
• Local backup and rollback workflow: each seat writes to its own directory under cha-backups\ with verify and restore commands
• Desktop interface built with Node.js; distributes as a Windows portable package
• Original dark manga aesthetic applied to documentation, workflow diagrams, and UI rather than a standard developer palette

Caveats

• The README never defines what “jailbreak” or “armor-breaking” means technically, or how the tool bypasses model safeguards beyond standard prompting
• Several listed model versions do not exist as public offerings; the README does not clarify if these are speculative targets, typos, or fictional set dressing
• The tooling appears Windows-first: examples use PowerShell, paths use backslashes, and the portable build targets Windows only

Verdict

A niche curiosity for prompt engineers and red-teamers who want a multi-model wrapper with strong visual branding. Most developers will find the fictional model roster and vague technical claims hard to take seriously.