Skip to content
Core-X Docs

Core-X Documentation

Core-X is a modular AI ecosystem designed for Apple Silicon Macs. It runs a constellation of MLX-accelerated services — LLM, RAG, vision, audio, embeddings, image generation, video generation, and speech-driven animation — all locally, with no data leaving the machine. A unified React + Three.js interface adapts to the active context (chat, canvas, voice avatar, image workflow, code, research, etc.), communicating with services through a central gateway and an SSE-based event bus.

The system is built around independent houses (domain-specific workspaces) that share infrastructure but maintain creative autonomy, an anthology of documentation and research, and a model zoo that tracks every model artifact used across the ecosystem.

Design Principles

  1. Local-first, private by default — All inference on Apple Silicon via MLX. No cloud APIs, no telemetry, no data exfiltration.
  2. Configuration-driven — 70 JSON schemas govern every entity. Services, agents, flows, skills, and houses are all declared, never ad-hoc.
  3. Tier-graduated resource management — Start with 3 core services on 16 GB; scale to 10+ services on 64 GB+.
  4. OpenResponses as canonical protocol — Every LLM-capable service speaks POST /v1/responses with SSE streaming. Legacy /v1/chat/completions is removed.
  5. House modularity — 8 creative houses operate independently on shared infra. Each registers capabilities via the core registry system.

Quick Facts

  • Core-X is a local-first, privacy-preserving AI platform running entirely on Apple Silicon via MLX
  • All LLM inference, vision, audio, embeddings, and animation happen on-device — zero external API calls
  • Services are organized into tiers (core / standard / full) that scale with your RAM budget
  • The Unified UI (React + Three.js) adapts across 16 modes (chat, canvas, voice, image, video, etc.)
  • The canonical LLM protocol is OpenResponses (POST /v1/responses) — not /v1/chat/completions