VibeTags operates on a straightforward premise: if you cannot trust an AI coding agent to avoid your critical infrastructure, annotate it out of reach. The Java annotation processor targets tools that operate across an entire repository — Claude, Gemini, Cursor, Codex CLI — and lets developers mark specific code regions as off-limits or as requiring human approval before AI-generated changes are applied. This is a structural acknowledgment that foundation model providers and enterprise engineering teams have different risk tolerances for the same codebase.
The timing matters. Developer trust in AI tool boundaries has eroded alongside a documented pattern of MCP infrastructure failures. The VS Code bug that shows tool_search returning between one and six tools from a pool of thirty-one — with no correlation to prior usage or semantic relevance — establishes the failure mode that makes annotation-level controls compelling. VibeTags does not fix the MCP reliability problem; it makes that problem less catastrophic. The developers annotating their codebases today are writing the governance layer that