🚀 New Features
Agent mode — an agent made for local and open models
- Full Skills support, so you can personalize the agent for your own tasks (#118)
- Web search powered by Exa, so the agent can go online
- Runs locally on llama.cpp with any GGUF model
- Reads, writes and edits files and folders on your device
- Approvals let you confirm each action manually or give the agent full access up front
- Fully local and private, works offline and nothing ever leaves your machine
🔧 Improvements & Fixes
- Repinned the 26B MTP assistant GGUF so 26B auto-downloads use the regenerated format
- Fixed MLX downloads freezing at 0% for models with a dot in the name like Qwen3.5 (#184)
- Failed downloads now show a toast instead of failing silently
- Better MLX checkpoint loading with correct text-only classification, a dedicated load mutex, and more accurate vision detection
- Windows no longer flashes console windows during backend version checks
- Tool call previews are now readable and scroll along as the agent writes (#188).
🙏 Contributors
Thanks to @Vect0rM, @worthant, @dtorey-d, @cristianorevil, @cashcon57 and @aryapratham1601-sketch for their contributions to this release!