Inference residency is not action residency.
A laptop running open weights through Ollama or llama.cpp can keep prompts and completions on machines you control. That is sampling residency. It is not the same claim as “nothing leaves this box.”
Once tools, web search/fetch, or MCP servers join the loop, the model still proposes — the host disposes. Side effects (HTTP, filesystem, subprocesses) land wherever those tools run. Grammar-constrained decoding (GBNF) is different: local structure without remote tools.
Protocol literacy — not product coaching.
#LocalAI #MCP #Ollama #TrustBoundary #DeAI

Inference Residency Is Not Action Residency: Tools, MCP, and Local LLMs
Local open-weight sampling can keep prompts on machines you control — that is inference residency. Tools, Ollama web search/fetch, and MCP server...


