Show HN: ChatGPT clone using open-source LLMs (github.com)

A new Kdeps example repository bundles a ready-to-deploy, offline-first “ChatGPT clone” (Dockerized, FE at localhost:8080) plus a set of APIs that demonstrate practical, on‑prem LLM applications. Highlights include an autonomous tool-selection API (MCP/A2A style) wired to two Python tools and running with llama3.2 (/api/v1/tools), a vision endpoint using moondream:1.8b that accepts image uploads (/api/v1/vision?q=...), and an OCR pipeline that combines Tesseract and Poppler-utils to return structured JSON (/api/v1/ocr). There’s also a weather “forecaster” agent that queries Open‑Meteo and synthesizes human-friendly advice (/api/v1/forecaster). Example curl outputs in the repo show complete JSON responses and sample extraction formats. For the AI/ML community this is significant because it stitches together open-source LLMs, multimodal models, classic OCR tools, and external APIs into deployable, privacy-focused demos—useful for on‑prem demos, rapid prototyping, and studying autonomous tool invocation patterns. Technical takeaways: the stack is Dockerized for local/offline use, exposes REST endpoints for tools/vision/ocr/forecast, uses llama3.2 for tool orchestration and moondream:1.8b for vision, and warns that structured JSONResponseKeys may require stronger models (e.g., llama). It’s a practical reference for building end‑to‑end, multimodal LLM applications with local control and reproducible examples.