ScrapChat ships self-hosted Qwen3.5 assistant

ScrapChat feels less like a chatbot and more like a local operator's cockpit. That's the right direction for local AI, but it also means the project wins or loses on setup friction and trust, not just model quality.

• –The stack is intentionally low-friction for tinkerers: Node/Express, vanilla JS, and Tailwind, with no bundler, keep the repo easy to modify.
• –The feature mix is unusually broad for a local assistant, with web search, Python, shell access, interactive charts, travel, and E*TRADE workflows all sharing one UI.
• –Sessions, prompts, and templates are the quiet killer feature, because repeatable workflows matter more than novelty once the assistant becomes a daily tool.
• –The 5090/Qwen3.5 combo targets serious consumer GPUs and local-model enthusiasts, so the audience is narrower than the feature list suggests.
• –The finance integration is the most ambitious part, but it also raises the bar for correctness and safety far above a normal chat app.