Per-Run Hard-Stop Token Budget Layer For Indie AI Coding Agent Subscriptions That Caps Catastrophic Loops Before They Eat the Monthly Quota
Solo developers on Cursor Max and Claude Code Max plans report single agent runs eating 79% of their monthly quota in 90 minutes (Anthropic confirmed deliberate weekday-peak rate-limit tightening on 2026-03-26), with one Max 20x user watching usage jump 21% to 100% on a SINGLE prompt. The unmet need is a session-level fuse box: set a per-run hard cap of $X or N tokens or M minutes, hook into the Cursor/Claude Code/Aider process, and kill the run automatically before a runaway loop wipes out the rest of the month.
Distinct from the published 4/28 'Agent-DB Safety Gateway' — that's about prod DB writes. This is about the indie dev's $200/mo subscription getting nuked by ONE bad recursion. Build it as a Cursor/Claude Code hook or MCP that aborts on cumulative iteration count, not after-the-fact analytics. Ship before Anthropic/Cursor add it natively, because they will.
landscape (3 existing solutions)
Anthropic and Cursor confirmed in March 2026 that limits tightened on purpose and there's no roadmap for hard per-run caps. A third-party MCP/extension that intercepts agent loops and enforces user-defined fuses is a clean unaddressed niche.