- 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.