Pre-flight budget gate for AI agent runaways. Drop-in. ~$0.001 per check.

One line, zero refactor

pip install burnstop

import burnstop
burnstop.install(budget_envelope_usd=5.00)

# your existing anthropic.Client calls work unchanged
# raises BudgetExceeded before token-1 if envelope would be exceeded

Why this exists

Date	Stack	Loss
Jul 2025	Replit AI `DROP DATABASE`	prod DB + 1,206 records
Nov 2025	LangChain A2A 11-day loop	$47,000
2026	LangGraph weekend refactor	$4,200
Apr 2026	Cursor + Claude → PocketOS	DB + 30h outage

Same shape every time: no per-run budget that lives outside the model's context. burnstop is that primitive.

What you see

[burnstop] OK call=1 proj=$0.0442 left=$5.0000
[burnstop] CALL#1 actual=$0.0435 cache_read=0% cumul=$0.0435
[burnstop] OK call=2 proj=$0.0442 left=$4.9575
...
[burnstop] 402 BUDGET_EXCEEDED: would_spend=$0.0442 envelope_left=$0.0080

Also exposes cache_creation vs cache_read decomposition Anthropic response headers don't surface (the GH cookbook #46829 issue).

Ecosystem adapters

burnstop-langchain — LangChain CallbackHandler · pip install burnstop-langchain
burnstop-langgraph — LangGraph node wrapper · pip install burnstop-langgraph
burnstop-crewai — CrewAI per-agent / per-crew gate · pip install burnstop-crewai

Stack

MIT licensed · 293 LOC · pytest passing · GitHub Actions CI on Python 3.10/3.11/3.12
Pure Python · stdlib + anthropic SDK (optional)
Drop-in monkey-patch · or manual Receipt() API
v0.1.0 — early. PRs welcome.

Built by an indie. github.com/santicalderon/burnstop · Hall of Pain · Agent Card