AI coding tools are eating your data and your budget
A wire-level analysis of xAI's Grok build CLI found it uploads the entire repository, every tracked file plus full git history, to xAI servers before the agent even starts working. The thread on that story was blunt: 'haha so they just stealing entire codebases?' Separately, a benchmark post found Claude Code sends 33k tokens to the model before reading the user's actual prompt, versus 7k for OpenCode.
The pattern here: AI coding agents are making aggressive, expensive, and sometimes invasive choices that users don't see. Claude Code's token bloat benefits Anthropic directly since more tokens means more revenue. Grok's wholesale upload benefits xAI's training pipelines. Both are rational from the vendor's perspective and problematic from the user's.
The counterpoint raised in comments is that bulk uploads let the model inspect code during 'thinking' without extra round trips, which could improve quality. But that's a performance tradeoff the user should be making consciously, not one baked in silently.
So what?
If you're running AI coding agents on a proprietary codebase, you need to audit what those tools are actually sending over the wire before you use them on anything sensitive. Token bloat also means your API costs for agentic tasks are likely higher than you estimated. Benchmark your chosen tool and compare alternatives like OpenCode before committing.