Augment Code Spurs Local Privacy Debate

The complaint lands: in AI coding, the model runtime is often the least sensitive part of the pipeline. The bigger privacy boundary is the context layer, and that’s where many "local" workflows quietly reintroduce cloud trust.

• –Augment's own docs say code is stored to power its context engine, so the criticism is about architecture, not trust.
• –The real privacy boundary is the whole context pipeline: retrieval, embeddings, code search, and telemetry can reveal more than a prompt ever would.
• –A local stack built from tree-sitter, SQLite, and on-device embeddings is technically feasible, but it has to beat hosted tools on ergonomics, latency, and recall quality.
• –The thread also shows the market split: some users want airtight privacy, while many local-model users mainly care about free inference and hardware control.
• –If privacy is the goal, teams should audit every external dependency in the context path, not just the model host.