Developer builds a local photo assistant with OpenClaw and MiniCPM-V 4.6

面壁智能 · wechat · 2026-07-28

Mohammed, an independent developer behind TechLatest, built a private photo assistant by wiring OpenClaw to MiniCPM-V 4.6 running locally via Ollama and exposing it through LitServe.

The workflow keeps images off external vision APIs: a custom vision-photo skill passes the local image path and user question to the model, which can OCR receipts, read whiteboards, inspect screenshots, and return structured results back into the same chat channel. A receipt demo returned the total $10.75 directly, showing a full loop from image intake to actionable reply. The article argues that on-device vision makes image understanding a reusable capability for agents, customer support, document processing, and field inspection.

Original post →

More from coding & agent

coding & agent channel →