feat: add MLX local model support for Apple Silicon

[wangericx]

Enables running LLMs locally on Apple Silicon (M1/M2/M3/M4) via the mlx-lm library, with no API key required.

• src/utils/mlx_lm.py: MLX inference engine wrapping mlx-lm generate, exposes a LangChain-compatible chat interface

• src/llm/mlx_models.json: curated list of MLX-compatible HuggingFace model IDs (Llama, Mistral, Gemma, Qwen, Phi families)

• src/llm/models.py: register 'mlx' as a provider; get_model() returns MLX chat model when provider is 'mlx'

• src/utils/llm.py: pass MLX_API_KEY / mlx base URL through LangChain call_options so the inference server URL is configurable

• app/backend/routes/mlx.py: GET /mlx/models endpoint returns available MLX models for the frontend model selector

• app/frontend/src/components/settings/models/mlx.tsx: UI panel to configure the MLX server URL and browse available models

• app/frontend/src/data/models.ts: fetch and merge MLX models into the global model list used by all node selectors

• docker/run.sh: pass MLX_API_KEY env var through to the container

  Summary

  • Run LLMs locally on Apple Silicon (M1/M2/M3/M4) via mlx-lm — no API key required
  • Add src/utils/mlx_lm.py: MLX inference engine with a LangChain-compatible chat interface
  • Add src/llm/mlx_models.json: curated list of MLX-compatible models (Llama, Mistral, Gemma, Qwen, Phi)
  • Register mlx as a provider in src/llm/models.py; get_model() returns MLX model when provider is mlx
  • Add GET /mlx/models backend endpoint to serve available models to the frontend
  • Add MLX settings panel in the web UI to configure the server URL and browse models
  • Frontend model selector automatically includes MLX models when the MLX server is reachable
  • Pass MLX_API_KEY env var through Docker

  Test plan

  • Install mlx-lm and start a local MLX server
  • Verify MLX models appear in the web UI model selector
  • Run a full analysis using an MLX model end-to-end
  • Verify non-Apple-Silicon machines gracefully skip MLX (no crash)