{"slug": "simple-self-hosted-llm-assistant-with-user-steered-compounding", "title": "Simple self-hosted LLM assistant with user-steered compounding", "summary": "A developer open-sourced Pawmc, a self-hosted LLM assistant that runs on Cloudflare Workers and Durable Objects, with user-steered compounding of context across sessions. The project costs $0 on Cloudflare's free tier, uses Workers AI as the default provider, and supports OpenRouter for larger models. The developer built it to learn Durable Objects and refined it after losing his job.", "body_md": "# Simple self-hosted LLM assistant with user-steered compounding\n\nI built a lightweight personal LLM assistant running on Cloudflare Workers + Durable Objects. The idea is simple: you pick a category and topic when starting a conversation, and the backend keeps a running summary per category/topic—so context compounds naturally across sessions without any\n\n[RAG](/en/tags/rag/)or embeddings overhead.No agentic magic, no vector stores. Just a clean singleton Durable Object that maintains state, Workers AI as the default provider, and optional OpenRouter support for bigger models. It runs free on Cloudflare's free tier with the default setup.\n\nI started this to learn Durable Objects and turn a rough idea into something usable. After losing my job recently, I spent more time refining it with friends' feedback—focusing on simplicity in deployment and daily use. This week I open-sourced it so others can try.\n\nRepo: https://github.com/kol3x/pawmc\n\nDetails:\n\n- $0 cost on Cloudflare free tier (default config)\n- Everything in a singleton Durable Object\n- Workers AI as the default provider\n- OpenRouter support for stable access and larger models (paid tokens)\n\nStory tracker · related coverage\n\n[LLM Routers: The Rise of a New Infrastructure Category 5d ago](/en/news/4436/)\n\n[Next AI Adoption Isn't Fixing R&D Waste — Here's Why →](/en/news/4984/)\n\n## All Replies （3）\n\nQ\n\nRan mine on a $5 DigitalOcean droplet—works surprisingly well for basic Q&A, though I had to trim the model size down to 7B to fit within memory.\n\n0\n\nS\n\nTried something similar on a Raspberry Pi 4—needed to offload context to disk after 2K tokens.\n\n0\n\nN\n\nCurious about the context window handling—do you truncate or summarize when conversations get long?\n\n0", "url": "https://wpnews.pro/news/simple-self-hosted-llm-assistant-with-user-steered-compounding", "canonical_source": "https://promptcube3.com/en/news/4986/", "published_at": "2026-08-04 19:28:32+00:00", "updated_at": "2026-08-04 19:50:34.019319+00:00", "lang": "en", "topics": ["large-language-models", "ai-infrastructure", "developer-tools"], "entities": ["Cloudflare Workers", "Durable Objects", "Workers AI", "OpenRouter", "Pawmc"], "alternates": {"html": "https://wpnews.pro/news/simple-self-hosted-llm-assistant-with-user-steered-compounding", "markdown": "https://wpnews.pro/news/simple-self-hosted-llm-assistant-with-user-steered-compounding.md", "text": "https://wpnews.pro/news/simple-self-hosted-llm-assistant-with-user-steered-compounding.txt", "jsonld": "https://wpnews.pro/news/simple-self-hosted-llm-assistant-with-user-steered-compounding.jsonld"}}