Local Sensei

Background Daemon taking screenshots at your command, sent to local VLM (hosted at your NPU/CPU/GPU), VLM-describe then send reflection to local coding LLM (hosted at your GPU).

Create successful code and automatically copy into copyboard <3

Requirements:

• Linux with X11/Wayland
• xbindkeys or other mapping key→command software
• Ollama
• at least 15 GB of free space (depends on selected LLM model)
• Recommended use case with Tilda or Yakuake (drop-down terminal)
• python3 with libs in requirements.txt

Example usage GPU (12GB VRAM):

Bind Client.py under some key with xbindkeys to trigger action

• Tried to do NPU+GPU combo but:
• [OpenVino NPU/CPU] Gemma 3 4b is unusable for recognizing code from IDE
• [OpenVino NPU/CPU] same as InternVL and Phi-3.5
• The best option is only to put GLM-OCR into VRAM, but it won't be at CPU/NPU (it's not compiled yet for OpenVINO for INTEL)

For LLM GPU:

• Ollama with Qwen3-8b (https://ollama.com/library/qwen3) ~ 6GB VRAM

For VLM GPU:

• Ollama with GLM-OCR (VisionLM) (https://ollama.com/library/glm-ocr) ~ 4GB VRAM