#mlx

Native macOS system cleaner with YAML-driven safety rules, local AI explainability via MLX, and an MCP server for agent-controlled cleanup workflows.

Local AI model manager

caix — native Apple Core AI inference server for Apple silicon (beta): OpenAI/Anthropic API, dashboard, streaming chat with tools/skills/MCP, MTP speculative decoding.

Local MLX tuned models through OpenAI compatible LLM, image generation and audio inference on Apple Silicon — Rust + PyO3 + MLX

A local multi-model gateway for Codex & Claude Code on Apple Silicon — a Qwable-first agent runtime (planner → executor → repair → critic → judge), fully offline.

Swift agent runtime for Apple platforms. Keeps tools, skills and history inside the 4,096-token on-device window, so connecting an MCP server doesn't break the turn. FoundationModels / MLX / OpenAI / Anthropic / Gemini.

Backend-agnostic black-box hallucination & RAG-faithfulness detection for LLMs — Probabilistic Certainty & Consistency (arXiv:2601.02574). Works on MLX / OpenAI / vLLM via token logprobs; no model internals, no training.

Mac-local LLM-from-scratch lab for building and evaluating a small decoder-only language model on Apple Silicon.

Local voice assistant for macOS Apple Silicon — wake-word → on-device STT → Claude → natural TTS, with speaker ID and DE/FR/EN.

Run Qwen3.8-27B fully offline on an Apple Silicon Mac (MLX, 4/5/8-bit) and use it as the local backend for Claude Code. Uncensored/abliterated build supported. No API key, no network. Built by yempik.

Thin, zero-dependency CLI to run local LLMs on Apple Silicon (vllm-mlx or mlx_lm) with OpenAI- and Anthropic-compatible endpoints — point Claude Code at a local model.

Fine-tuned 3B model + RAG beats Claude Opus 4.8 raw at entity resolution (99% vs 78.5%) on unseen companies — Covent LLM Challenge submission.

Your local and open source business AI toolkit for finetuning, analysis and...synthetic data, LoRA fine-tuning workflow, HDC retrieval, offline RAG

Delegate text work to a local model on oMLX from Claude Code, Codex, and other MCP clients — the content never leaves your machine

LoRA fine-tuning of Qwen2.5 (4-bit) with Apple MLX for freight-text→JSON extraction — honest before/after eval (71%→100% on held-out synthetic set), every number a committed receipt.