Other things in the box:

  - Ask-user tool. When the agent hits a judgment call (cost
    confirmation, ambiguous field, a captcha), it pauses and asks you
    in the chat, not in a page overlay. You answer, it resumes.
    I'm really obsessed with the ask user tool on Claude Code, 
    and obviously I implemented it here also.
  - Use oracle to plan complex tasks, take its help when stuck, and also to 
    create skills
  - Sessions. Each task is its own session with its own tab(s) and
    history. Switch between them and let the tasks run in the background.
  - Bring your own key. Gemini, OpenAI, Anthropic, or any
    OpenAI-compatible endpoint. No server of mine in the loop.
  - Skills. Teach it or let it figure out a reusable flow and save it as 
    skill to reuse it.
  - Auth handoff. When a login popup opens, the agent blocks, you
    complete the auth, the agent picks back up. I purposefully didn't automate
    things like auth/captcha, as the expectation of the current websites'
    implementations isn't automation.
  - Everything local. Traces of every run go to ~/.nimbus/traces/.
    No telemetry, YET. Nothing reaches my servers, you just contact the LLM 
    providers directly.

Looks interesting. But how does it differ from your typical AI browsers that exist? Did you find any usage differences as UX is changed?