Claude Code Local v1.0.0 came out today, November 3. It is built for one moment. You’re in the middle of something and Claude Code says your 5 hour limit is reached, and the reset is hours away, maybe days if it’s the weekly one. Type exit, then keep going, and the same conversation carries on with a model running on your own machine.
Since the first release in March it has grown from a Mac server into a one command install that runs on Mac, Linux and Windows, fits a 16 GB MacBook, and can hand the conversation over by itself when the limit hits. October shipped five releases (v0.4.0, v0.4.1, v0.5.0, v0.6.0 and v0.7.0) without much said about them. Version 1.0 puts them under one name.
curl -fsSL https://raw.githubusercontent.com/nicedreamzapp/claude-code-local/main/install.sh | bashThat line is for Mac and Linux. On Windows, paste this into PowerShell instead.
irm https://raw.githubusercontent.com/nicedreamzapp/claude-code-local/main/install.ps1 | iexOne command, and it picks the model
The installer checks the machine, picks a model that fits its memory, installs everything and puts Claude Local on your Desktop. Add --with all and it also brings in voice, Moonstone, the browser agent, phone control, remote screen, failover and the Trinidad Head terminal. Each add on installs through its own repo’s installer and has its own uninstaller, and your data stays unless you ask for it to be purged. A doctor script checks the core and every add on you installed, repairs what it can, and writes a privacy safe report for a bug report. There is a real version number now too: the server banner, the health check, both installers and both doctor reports say 1.0.0.
Mac, Linux, Windows, and small Macs
On a Mac nothing changes: MLX on Apple Silicon is still the engine. Off a Mac, the installer looks at your hardware (NVIDIA, AMD or Vulkan, or just a CPU), downloads a pinned llama.cpp build, and picks the strongest vision model that fits your memory at the highest precision it can. The llama.cpp build is pinned by version and checksum, and every model file by size and SHA-256.
A 16 GB Mac gets Qwen 3.5 4B at bf16, and an 8 GB Mac gets Qwen 3.5 2B at 8-bit. Both read images, and both passed the Claude Code test of reading a file, fixing a bug with Edit, running a script and describing an image. On a Mac that small, keep requests to one clear step at a time. The measured numbers are in SMALL-MACS.md.
The limit handoff
Type exit, then keep going, and your last Claude Code conversation in that folder carries on with a local model, or a free cloud model if you add an OpenRouter key. Picking a cloud model sends the conversation to OpenRouter; the local ones don’t leave the machine. There is also an automatic version: one hook opens a new window in the same folder when Claude Code stops on a rate limit, and the conversation carries on with no menu. Plain claude is never changed by the install, so it still uses your Claude plan when the limit resets. Both handoffs are Mac only for now.
Most of the other fixes since launch are the same kind of bug: something the model or Claude Code sent got quietly thrown away by the tool call parser. Those are fixed for Gemma 4 and Hermes 4, and the test suite passes 477 tests without needing a model loaded. v0.4.1 also closed a high severity security report in the handoff hook.
What it can’t do yet
A local model is not Claude Opus. Big models on a 64 or 96 GB Mac handle real multi step work. On 8 or 16 GB, keep to one clear step at a time, and a task that overflows the 16K context compacts and often loses the thread. Phone control and failover are Mac only, and there is no measured number for a base M1, M2 or M3 yet.
The full list is in the v1.0.0 release notes. The guide site has the usage limit page with the exact limit messages and how to switch to local and back, a page on running Claude Code on Apple Silicon with MLX, and one on which local model to run. The project page is here, and the code is on GitHub.
Thanks to everyone who filed issues and sent pull requests, including @zwolf25, @pratikgx, @tadrianonet, @KaoCSC and @anelyberdecias-oss.