Behind a gateway

Complements, not competitors

A gateway (LiteLLM, Portkey, Kong, …) sits between your app and the providers: it holds the org key, routes and falls back across models, and enforces a flat per-key budget. Breakerbox governs the internal structure of one run — a hierarchical dollar budget that trips at hop boundaries. Different layers, so run both: the gateway caps the org; breakerbox catches the runaway run under that ceiling.

Because breakerbox meters through a LangChain callback (not the network), it composes with any OpenAI-compatible gateway with no extra wiring — point your model at the gateway's base URL and wrap the app in guard().

LiteLLM

Your LiteLLM proxy holds the provider key and routes; the model just talks to the proxy.

from langchain_openai import ChatOpenAI
from breakerbox import guard

model = ChatOpenAI(
    model="openai/gpt-4o",              # the name breakerbox meters (match the price table)
    base_url="http://localhost:4000",   # your LiteLLM proxy
    api_key="sk-litellm-...",           # a LiteLLM virtual key, not a provider key
)
# build your LangGraph app with `model`, compile it, then:
app = guard(compiled_app, budget_usd=5.00)   # gateway routes + holds the key; guard budgets the run

Portkey

model = ChatOpenAI(
    model="openai/gpt-4o",
    base_url="https://api.portkey.ai/v1",
    default_headers={"x-portkey-api-key": "...", "x-portkey-virtual-key": "..."},
)
app = guard(compiled_app, budget_usd=5.00)

Model naming

Breakerbox meters by the model name it sees, so the name your gateway reports must match the price table (e.g. openai/gpt-4o). If the gateway surfaces a bare deployment name breakerbox doesn't recognize, it fails before the call is dispatched (never silently at $0) — add the model to the table, or pass unknown_model="default_rate" to meter it at a conservative default.