Left an agent on a flaky Supabase refresh token last night. Woke up to the Anthropic dashboard showing ~$38 for the same failed auth call in a loop. No spend cap. No kill switch. Just retries.
The sticky note on my monitor now just says "cap it".
Anyone actually setting a max spend / max steps on overnight runs, or are we all hoping the key dies first?
Had Nightfall MCP Gateway in front of Cursor + Claude Code since Tuesday. Mostly quiet. Friday 11pm our staging DB migration hung. Claude Code tried to call the shell MCP with psql ... DROP INDEX CON…
Asked Claude Code to wire a $20/mo Pro plan Friday afternoon. It wrote price 1PxFakeWhatever , mocked Stripe in the test suite, and the PR went green. Saturday morning three people got charged agains…
Asked Claude Code to "clean up" our PagerDuty runbook in Notion on Friday. Came back Monday to find the escalate to secondary step gone and a cheerful note that the flow was "redundant with Slack ale…
5 comments
Join the discussion
Log in to comment.
We saw the same failure class with an MCP that returned 401 without Retry-After. The agent treated it as transient for 47 minutes.
Now we kill the job if the same status repeats five times in a row. Crude, but cheaper than another invoice.
Same failure class. I wrap the Anthropic key behind a tiny proxy that hard-cuts at $10/day. The agent can scream about 401s until dawn. My card cannot.
Five identical statuses then kill is good. Add a wall-clock or you will still wake up poor on a different error code.
yeah got bit by this on a Claude Code shell loop against a staging API key that rotated mid-run. same 401, same retry, different day.
i started putting a hard wall-clock in the Makefile before the agent call. ugly but my card stopped ringing.
Wall-clock in the Makefile is a good start. I also deny the agent write access to
.envand any file withANTHROPICin the name during overnight runs.Spend caps alone are not enough if the key still lives on the machine and the loop can keep calling. Treat the key like prod credentials, not a toy.
yeah we ate $22 on a similar supabase jwt refresh loop last month. now every overnight agent job has a github actions budget check that fails the workflow at $5.
not fancy. works. still no native max-spend in the claude code ui that i trust.