- WEBUI_AUTH=true: require login (auth was disabled by default) - ENABLE_TOOL_SERVERS=true: unlocks the MCP/OpenAPI tool server UI https://claude.ai/code/session_012gDnantBmFTWZGCiKyjazx
NVIDIA dropped the distro-specific repo URL in favour of a single stable path. Also pass --yes to gpg --dearmor to avoid the interactive overwrite prompt on re-runs. https://claude.ai/code/session_012gDnantBmFTWZGCiKyjazx
Merges all components into one script — no separate server files needed. Embeds server.py (RAG) and mcp_server.py (MCP) as inline heredocs. - Auto-detects VRAM and sets models + context window accordingly: 6GB → 7B models, 8k ctx 8GB+ → 14B chat + 7B code, 16k ctx 16GB → 14B models, 32k ctx (V100/A100 friendly) - New install or update (detects existing compose) - --force flag to overwrite config files - --no-pull to skip model download prompt - Adds fetch_url tool to MCP server - Optional DeepSeek-R1:14b pull for reasoning tasks https://claude.ai/code/session_012gDnantBmFTWZGCiKyjazx
Claude