Local AI Telemetry Report 2026: which local AI tools phone home, and what they send
Abstract
Local model does not mean local software. This first-edition report from Stridenalysis, the research desk of StrideNote, brings together eleven investigations published between June and August 2026 on one question: what does a local AI tool send from the machine it runs on? Seven tools from a working local AI stack were installed on one test machine (Apple M4 Pro, 48 GB, macOS), their code read, their network connections listed and, where possible, their traffic captured: OpenJarvis, OpenCode, Pi, LM Studio, huggingface_hub, Graphify and Visual Studio Code. Every finding carries an evidence grade (A capture, B sockets, C code). Four of 13 findings show data sent to a reporting service without the user asking; no capture showed a prompt or file leaving the machine. The report records six false leads, a Claude Code case reported by others, five checks anyone can run, its limitations, and the tests planned for the next edition. Files: the report (PDF, 30 pages) and the dataset behind every chart (local-ai-telemetry-2026.csv, one row per finding).
