Financial circuit breaker for AI agents

$4,200

One agent. One loop. One weekend. No cap.

Set a cap.
Never see a bill
like this.

A drop-in LLM proxy with a hard, atomic per-run budget cap. Set $0.10 or $50: the first call that would exceed it gets a 402. The loop dies. The bill stops. No code changes required.

Start free → See the code ↓

1,000 runs / month free • no credit card required.

agent.py · loop-001

▋

Live demo

See the cap enforce. Live.

Live calls to the Anthropic API, real enforcement. Not a staged demo. A rogue agent hits the cap and dies, then 50 parallel sub-agents share one budget with zero overspend.

Burnbreaker is a safety tool, not a guarantee.

Integration

Your existing SDK.
Two new headers.

Point your Anthropic or OpenAI client at the Burnbreaker endpoint. Add X-Gateway-Key and X-Budget-Cap. Every call on that run is pre-checked before the request reaches the provider.

No SDK wrappers. No instrumentation. No infrastructure to manage. If you can set an HTTP header, you’re protected.

anthropic

openai

import anthropic client = anthropic.Anthropic( base_url="https://burnbreaker.com", api_key="sk-ant-...", # forwarded, never stored ) message = client.messages.create( model="claude-sonnet-4-6", max_tokens=1024, messages=[{"role": "user", "content": "..."}], extra_headers={ "X-Gateway-Key": "gw_...", # from API Keys page "X-Budget-Cap": "0.10", # hard cap in USD "X-Run-Id": "agent-run-001",# unique per run }, )

Shared budget • atomic reserve

AGENT #01

$0.18 reserved

AGENT #02

$0.18 reserved

AGENT #03

$0.18 reserved

AGENT #04

$0.18 reserved

AGENT #05

$0.18 reserved

AGENT #06

$0.18 × BLOCKED

Shared budget • $1.00

committed$0.72

reserved$0.18

remaining$0.10

agent #06 needs$0.18, blocked

Concurrency safety

50 sub-agents.
One atomic budget.
Zero overspend.

When your orchestrator fans out to 50 parallel agents, a naive counter lets them all read “budget available” at the same instant and collectively spend 50× what you allowed.

Burnbreaker uses an atomic Redis Lua script to reserve budget before the provider call, not tally after. The first agent to exhaust the cap is blocked; all others proceed within their reserved slices. The cap is a hard ceiling, not a guideline.

Burnbreaker is a safety tool, not a guarantee.

The race condition

“Just track spend in Redis” doesn’t work.

A concurrent read-then-write is a TOCTOU race. All agents read the same balance, all pass the check, all commit, leaving you far over cap before any single call completes. Burnbreaker closes the gap with a single atomic operation: check and reserve happen as one indivisible step.

✗ Naive: TOCTOU race

# All 50 agents read $0.00 at t=0 spent = await redis.get(run_id) # $0.00 cost = estimate(request) # $0.50 if spent + cost > cap: # 0.00+0.50 < 1.00 ✓ raise BudgetError() # ← 49 others pass the same check await redis.incrbyfloat(run_id, cost) # 50 commits → actual spend: $25.00 🔥

✓ Burnbreaker: atomic Lua

# Reserve and check are one operation SCRIPT = """ local s = tonumber( redis.call('GET', KEYS[1]) or 0) local c, cap = ARGV[1], ARGV[2] if s + c > cap then return -1 end return redis.call( 'INCRBYFLOAT', KEYS[1], c) """ result = await redis.eval( SCRIPT, 1, run_id, cost, cap) if result == -1: raise BudgetExceeded() # 402 pre-call

Pro: Smart Control

When budget pressure builds,
downgrade. Don’t die.

Pro accounts get three additional controls that fire transparently between the gateway and your provider. No agent code changes required.

Model Downgrade

Once committed spend crosses a threshold (e.g. 80% of cap), Burnbreaker silently rewrites the request to a cheaper fallback model. The run continues at reduced cost instead of hitting a hard stop.

X-Fallback-Model: claude-haiku-4-5-20251001 X-Downgrade-Threshold: 0.8

CALLS 1–16 claude-sonnet-4-6 $0.018/call

── 80% of budget reached ──

CALLS 17+ claude-haiku-4-5 $0.0008/call

Kill Rules

Set a maximum call count or error streak. If either threshold is crossed, Burnbreaker returns 402 before the next call fires, stopping runaway loops and retry spirals before they compound.

X-Max-Calls: 20 X-Max-Errors: 3

Webhook Alerts

Burnbreaker POSTs a structured JSON payload to your endpoint whenever a run hits its cap, a model downgrade fires, or a kill rule triggers. Wire it into PagerDuty, Slack, or your own pipeline.

X-Alert-Webhook: https://your.app/hook

Pricing

Simple. No usage fees.

Flat monthly rate. Your LLM costs go directly to Anthropic or OpenAI. Burnbreaker never touches or marks up what you spend with providers.

Free

/ month

Get started

Hard atomic budget cap per run
Concurrency-safe enforcement
Run tracking & spend dashboard
Anthropic & OpenAI support
1,000 runs / month
Model downgrade
Kill rules
Webhook alerts

Pro

$29

/ month

Start Pro

Everything in Free
Unlimited runs / month
Model downgrade on budget pressure
Custom kill rules (calls & error streak)
Webhook alerts on any event
Priority support

Your token costs flow directly to your Anthropic or OpenAI account. Burnbreaker only meters usage and enforces caps. We never route, hold, or mark up what you spend with providers.

Burnbreaker is a safety tool, not a guarantee. You remain responsible for monitoring your own provider spend.

Set a cap.Never see a billlike this.

See the cap enforce. Live.

Your existing SDK.Two new headers.

50 sub-agents.One atomic budget.Zero overspend.

“Just track spend in Redis” doesn’t work.

When budget pressure builds,downgrade. Don’t die.

Simple. No usage fees.

Set a cap.
Never see a bill
like this.

Your existing SDK.
Two new headers.

50 sub-agents.
One atomic budget.
Zero overspend.

When budget pressure builds,
downgrade. Don’t die.