Home › Free tools › Agent Runaway Check

Agent Runaway Check: is your coding agent one setting away from burning your weekly limit?

A free, one-minute check for Claude Code and Cursor. It finds the settings that let agents run unattended, launch more agents and loop, and shows which of your sessions used the most. One Python file, read-only, MIT licence.

It runs on your machine, not in this page. The script reads your Claude Code and Cursor settings and your local Claude Code transcripts. It changes nothing, sends nothing and needs no account. Python 3.9+, standard library only.

Download runaway_check.py (free) MIT licence

v1.0.0 · sha256 7e5f743495daf18954e95f027040d5d8b20ab1854eed897de099b3c08c98123b

python3 runaway_check.py                       # your home folder + the current project, last 14 days
python3 runaway_check.py --project ~/code/app --days 30
python3 runaway_check.py --json

Exit code 1 if anything HIGH is found, so it also works in CI.

What it checks

HIGH: Claude Code bypassPermissions mode; an allow rule that lets the agent start another agent (Bash(claude -p:*), Bash(codex …)); Cursor CLI approvalMode: "unrestricted" or a Cursor allow rule for Shell(cursor-agent); an alias that always passes --dangerously-skip-permissions.
MEDIUM: blanket Bash or Shell(*) permission; an alias that always runs cursor-agent --force; past sessions with FAN-OUT (8+ subagents) or LOOP (the same tool call 5+ times); settings files that don't parse.
INFO: no Claude Code or Cursor hooks, so nothing can stop a runaway session.
Your 5 heaviest Claude Code sessions in the period, each API response counted once. "Weighted tokens" is a relative cost measure, not your plan's meter.

How it was tested

14 automated tests against fake home folders in the documented formats, plus a run on 5 transcripts written by the real Claude Code 2.1.314 CLI (test sessions with a scripted model, so not real-world usage) and checks of the settings paths and flags against Claude Code 2.1.314 and cursor-agent 2026.10.22. Not tested: a live Cursor session, transcripts from real paid use, macOS and Windows. Read its TEST-REPORT.

Want it enforced, not just reported?

Agent Spend Guard (US$39) caps subagents, blocks nested agents and retry loops, and stops sessions at your token budget. Its 49 automated tests include 6 hook runs inside the real Claude Code CLI; it has not had a live Cursor run, and in Cursor only its nested-agent and repeat-call rules apply.

See Agent Spend Guard · Read its test report