llm-circuit-breaker

A tiny, readable circuit breaker for LLM calls. When your budget runs out, it doesn't raise an exception or let the bill run — it falls through to a free local model so your agent keeps working.

Why

Agent API costs are unpredictable: most runs cost a few cents, but a stuck loop or an unbounded retry can burn tens of dollars in minutes. Most LLM gateways solve observability (dashboards, logs) or routing (pick the cheapest/best model) — but none of the popular ones hard-stop into free local compute the moment you're about to overspend.

This library does one thing: try providers in order, and once you've hit your budget, skip straight to a local model instead of failing or overspending. No hosted proxy, no dashboard, no vendor lock-in — about 150 lines you can read in five minutes and vendor straight into your own project.

Install

pip install llm-circuit-breaker
# or, with Anthropic support:
pip install "llm-circuit-breaker[anthropic]"

Quickstart

from llm_circuit_breaker import LLMCircuitBreaker, anthropic_tier, ollama_tier

breaker = LLMCircuitBreaker(
    tiers=[
        anthropic_tier(api_key="sk-ant-...", model="claude-sonnet-4-5"),
        ollama_tier(model="gemma:2b"),  # local, free, always-on fallback
    ],
    daily_limit_usd=5.0,
    ledger_path="llm_spend.jsonl",
)

result = breaker.complete(system="You are terse.", user="2+2?")
print(result.text, result.tier_used, result.cost_usd)

Once llm_spend.jsonl shows $5 spent today, every subsequent call for the rest of the day skips Anthropic entirely and goes straight to your local Ollama model — automatically, no code change, no surprise bill.

v0.2 — streaming, async, and a cost CLI

Stream tokens as they arrive (complete_stream) — same tiers, same budget fallback, just yielded piece by piece. Built-in Anthropic / OpenAI / Ollama tiers stream natively; any tier without streaming yields its whole result once.

for chunk in breaker.complete_stream(system="You are terse.", user="Explain RAG in 2 lines"):
    print(chunk, end="", flush=True)

Async (acomplete) — keep many calls in flight at once:

import asyncio
results = asyncio.run(asyncio.gather(*[
    breaker.acomplete(system="s", user=q) for q in questions
]))

See where the money went — a budget breaker you can actually inspect:

llm-cb cost                 # today / total / per-tier, straight from the ledger
llm-cb cost --ledger path   # point at a different ledger, or --json for raw

LLM spend  (llm_spend.jsonl)
   today : $2.3100
   total : $45.8700   (1240 calls)
   by tier:
     anthropic:claude-sonnet-4-5      980 calls   $44.2000
     ollama:gemma:2b                  260 calls   $0.0000

How it decides

1. Check today's + all-time spend against your limits (a plain JSONL ledger file — human-readable, no database).
2. Under budget: try each tier in order, return the first success.
3. Over budget: skip every paid tier and try only the tiers marked is_local=True.
4. If every attempted tier fails (or the budget is blown with no local tier configured), raise a clear exception instead of failing silently.

Built-in tiers

anthropic_tier, openai_tier, gemini_tier, openrouter_tier, and ollama_tier (local, via Ollama) — or write your own in about 5 lines; a Tier is just a name and a (system, user, max_tokens) -> (text, cost_usd) function:

from llm_circuit_breaker import Tier

def _call(system, user, max_tokens):
    text = my_provider_sdk.chat(system, user, max_tokens)
    return text, my_own_cost_estimate

my_tier = Tier(name="my-provider", call=_call, is_local=False)

What this is not

Not a hosted gateway, not a dashboard, not a replacement for LiteLLM or Portkey if you need enterprise routing, guardrails, or a proxy server in front of a team's traffic. This is the opposite trade-off: minimal, in-process, one job — for solo builders and small agent projects who want a hard stop before overspend without standing up another service.

Pricing table

pricing.py ships an approximate per-model $/1M-token table used only to estimate spend for the budget check. Prices change and vary by region/tier; override PRICING["your-model"] = {"input": ..., "output": ...} for accuracy, or return your own cost from a custom tier.

Development

pip install -e ".[dev,anthropic]"
pytest

License

MIT