0.0.28 — Claude Code on your local model, and an honest pipeline

Highlights since 0.0.23:

• Claude Code runs on your local model: the filtered endpoint speaks the Anthropic Messages API (/v1/messages, streaming and tool calls included). The Clients page wires claudeCode.environmentVariables into VS Code/Cursor/Insiders/VSCodium — with a timestamped backup, a one-click Remove, and a terminal snippet for one-off sessions.
• One owner per surface: with the app installed the extension is chat-only; the app owns all panels and settings.
• Model residency you can trust: API-initiated loads update the dashboard, launches verify instead of trusting caches, stops verify their kills, and the GUI follows model switches within seconds.
• Downloads and conversion overhauled: only MLX-usable files are fetched (gpt-oss-120b: ~65 GB instead of 195 GB), truthful progress, orphaned workers reaped, cancel/retry races gone, dismissible cards, and Search knows what's already on disk.

Full details in CHANGELOG.md. macOS Apple Silicon; unsigned builds — right-click → Open on first launch.

MLX Console GUI 0.0.23

First public release. macOS (Apple Silicon) desktop app and VS Code extension for running large language models locally through mlx-lm — with honest memory accounting before, during and after.

Install:

• Desktop app (primary): download the DMG below, open it, drag to Applications. Unsigned build — right-click → Open on first launch.
• VS Code extension: download the VSIX, then code --install-extension mlx-console-gui-0.0.23.vsix.

Highlights of this build: the Info page as tabbed docs with rendered diagrams, Hugging Face search with exact sizes for every quantization family, conversion with bf16 and 5-bit options and up-front refusals for what mlx-lm cannot run, resumable downloads that survive kills and restarts, staged settings with an explicit Save, and the full API documented — including what is not on it.

See CHANGELOG.md for the detailed history.