Claude Code Local v0.5.0 支援 Linux 和 Windows，v0.6.0 加入配套工具安裝

Back in August someone opened issue #53 on claude-code-local asking for a Windows version. They had an HP laptop with 32 GB of RAM, and the installer stopped them because it only ran on Apple Silicon. I told them the truth at the time. The whole server is MLX native, MLX doesn't exist on Windows, and their RAM was plenty. The chip was the blocker. I said if anyone ported it to llama.cpp I would link it, and nine days later I closed the issue as Mac only.

On October 29 I went back to that thread with a different answer. Claude Code Local runs on Windows now, and on Linux too.

If you haven't seen the project, claude-code-local lets you run Claude Code against an AI model on your own computer instead of the cloud. Your code and your prompts stay on your machine. It's at a little over 3,390 stars on GitHub, and for most of its life it was a Mac thing, because MLX on Apple Silicon is what I run every day. That's still the heart of it.

What changed is the part in the middle. The server that speaks the Anthropic API and translates tool calls used to be tied to MLX. Now that translation doesn't care which engine sits underneath, so MLX and llama.cpp share one code path. On a Mac nothing changes. Off a Mac, the installer checks whether you have an NVIDIA card, an AMD or Vulkan GPU, or just a CPU, downloads a pinned llama.cpp build for it, and picks the strongest vision model that fits your memory at the highest precision it can. A 24 GB NVIDIA card gets Qwen 3.5 9B at full BF16. A 12 GB card gets Qwen 3.5 4B at full BF16. With no GPU at all it sizes a 4B or 9B vision model to your RAM. Every model it can pick reads images, so a screenshot you paste into Claude Code works the way it does in the cloud.

On Windows it's one line in PowerShell, and then you double click Claude Local on your desktop. On Linux it's one line in a terminal and you run claude-local. That was v0.5.0, and I called it Linux and Windows. I posted the PowerShell line in issue #53 and said that if anything breaks on their machine, open a new issue and I'll jump on it.

Less than an hour later v0.6.0 went out, and that's the one I'm actually excited about. I've built a lot of pieces around Claude Code, each in its own repo, and until now they were separate things you had to find and set up yourself. The local model is the brain. The rest give it ears, a voice, hands and reach. So now the installer offers them.

curl -fsSL https://raw.githubusercontent.com/nicedreamzapp/claude-code-local/main/install.sh | bash -s -- --with all

Run it without --with and it shows you a menu once the model is set up. On Windows you set $env:CCL_WITH = "all" before the PowerShell line to skip the menu and get everything.

Here's what's in the menu. Voice is NarrateClaude, so you can talk to Claude Code and hear it answer. Moonstone puts every session on every machine in one window. Browser is the browser-agent broker, which lets agents use your real browser with each one in its own tab. Phone is claude-screen-to-phone, for driving Claude Code from your iPhone. Remote is FiaOS, which gives you your screen and a real terminal from any browser. Failover is claude-failover, which keeps your agents working on a local model when you hit your limit. Terminal is Trinidad Head, a terminal I wrote from scratch in Rust that gives every window its own glow color.

Each one installs through its own repo's installer, so every piece still works on its own if that's all you want. Every repo got an uninstaller too, and removing everything leaves your data alone unless you add the purge flag. A couple of them got smarter about each other along the way. claude-screen-to-phone uses the browser broker now instead of launching its own browser, and claude-failover uses the claude-code-local server when it's already running instead of loading a second model.

The voice piece got the biggest rewrite. NarrateClaude used to lean on Apple's speech recognition, which is why it only worked on a Mac. Now it hears you with whisper.cpp running Whisper large-v3-turbo at full f16 precision, and it answers with a local Piper voice. Instead of typing into Apple Terminal, it runs claude inside a pty the loop owns (ConPTY on Windows), so it works in any terminal. That's what let it go to Linux and Windows the same day. Apple's engine is still there behind a flag if you want it, and everything in the voice loop is open source and runs offline.

So whoever was on that HP laptop can have the model now, and they can also talk to it, watch every session from one window, and hand it their browser. The repo is github.com/nicedreamzapp/claude-code-local. The README already says what's next. Phones are the next frontier, and we're working on it.