Overview

flowchart LR
    A[Telegram / WhatsApp / CLI] -->|photo + caption| B[OpenClaw Gateway]
    B -->|minicpm-v4.6 plans| C[vision-photo skill]
    C -->|vision_query.sh| D[LitServe Vision API :8002]
    D --> E[(Ollama MiniCPM-V 4.6)]
    E --> D
    D --> C
    C --> B
    B -->|structured reply| A

1. User sends a photo on Telegram, WhatsApp, or CLI
2. MiniCPM-V 4.6 handles chat and invokes the vision-photo skill
3. Skill POSTs to LitServe http://127.0.0.1:8002/predict
4. MiniCPM-V analyzes the image locally — OCR, summary, suggested channel reply
5. Answer returns through OpenClaw to the same channel

Layer	Role
OpenClaw	Channels, sessions, skills, daemon
minicpm-v4.6	1.3B vision model — chat + image input (~1.6 GB)
vision-photo skill	Shells out to vision_query.sh
vision_server.py	LitServe API wrapping Ollama vision calls

Animated workflow

Guide	Overlap
MiniCPM-V MCP Server	Same model as MCP tools in Cursor
OpenClaw + Gemma + RAG	Text RAG skill pattern (this guide adds photos)
MiniCPM-V Benchmark	vs Qwen3.5-0.8B and Gemma4-E2B

Full tutorial

See TUTORIAL.md.

Read the full tutorial →