Show HN: Julie update – local LLMs, CUA, installers and perf gains

[luthiraabeykoon]

The biggest shift is that Julie now supports fully local LLMs and agentic workflows. It’s no longer limited to answering questions about what’s on screen. It can now run writing and coding agents, and optionally take concrete actions on your computer under supervision.

What’s new:

- Local LLM support. Julie can now run entirely on-device,

- Agentic computer use. I added a computer-use mode with demos showing multi-step actions like clicking, typing, and navigation.

- Writing and coding agents. Draft, refactor, and iterate in-place without moving into a separate workspace.

- Installers are now available, so setup is a lot simpler.

- Significant performance improvements across startup time, memory usage, and latency.

I also wrote a full walkthrough and demos covering how the agents work and where the boundaries are: https://tryjulie.vercel.app/

Repo + installers: https://github.com/Luthiraa/julie

Thanks for all the support and feedback. From the bottom of my heart, I really appreciate it. I’ve loved building this, and it’s been one of the fastest things I’ve taken from idea to something real that people are actually using. I really love this community.

If you enjoyed checking it out, a star on the repo would mean a lot and helps more people find it.

[llamatheollama]

This is a really thoughtful direction. The “overlay instead of another workspace” idea resonates a lot, especially the screen-as-context inversion.

Curious about one thing: where local LLMs feel “good enough” today vs where you still fall back to remote models

The perf work + installers make this feel way more real than most agent demos. Nice job shipping something people can actually try.

[luthiraabeykoon]

Good question. Local LLMs are already “good enough” for most in-context work: short-to-medium writing, refactors, reasoning over what’s on screen, and multi-step agent plans where latency and privacy matter more than raw IQ.

I still fall back to remote models for very long-context tasks, heavier code synthesis, or when you want best possible reasoning over large codebases, the goal is to default local, then escalate only when it actually adds value.

[luthiraabeykoon]

One small note: I’m finally at a place where I’m genuinely happy with where this landed, so I’ll probably pause active development for a bit.

That said, I’m excited to see how people use it, and I’ll still be around to answer questions and fix issues if they come up.

[moderngamer]

what llm do you recommend using locally?

[luthiraabeykoon]

i use qwen3-8B for text and qwen3-vl-4B for vision