Kintsugi – AI Agent Safety Net

AI coding agents can execute dangerous shell commands (rm -rf, DROP TABLE) without safeguards. Kintsugi acts as a local safety layer that intercepts destructive commands before execution, explains the risk in plain language, and makes dangerous actions reversible with snapshots—preventing catastrophic mistakes.

Agentic coding tools—Cursor, Devin, Claude Code—are moving fast and the "oops the agent deleted everything" failure mode is already a meme with real casualties, which means the pain is documented and growing. No clear incumbent owns this safety layer space yet, though some agent frameworks are starting to bolt on permissions natively, which is the existential threat here. The revenue band is honest: this is a prosumer and small-team tool that works as a low-cost local utility, so $2k–8k/month is achievable through volume but unlikely to stretch much further without an enterprise angle. The biggest risk is that the major agent platforms—Anthropic, Cursor, OpenAI—treat this as a missing feature rather than a missing product and ship it themselves within 12 months, leaving no moat.