Introduction
Breakerbox is the circuit breaker for AI agents — a hard dollar budget on the workflow itself.
What is Breakerbox?
Breakerbox puts a hard dollar budget on an agent workflow itself — hierarchical across
sub-agents, enforced between steps, and never a mid-flight failure. The Python package is
breakerbox.
Dashboards are autopsies. A budget that trips between steps is a seatbelt.
Most cost tooling tells you what an agent spent after it spent it. Breakerbox stops the run at a hop boundary the moment it would exceed the budget you set — so your agents can't outspend you.
The two pieces
- A library.
guard()wraps a compiled LangGraph app and enforces a hierarchical dollar budget (reserve → execute → reconcile, like a credit-card hold). It self-meters tokens, trips gracefully at hop boundaries, and emits a receipt. - A visual builder. A canvas that generates readable, guarded Python you run yourself — codegen only, never server execution.
What Breakerbox will never do
These are permanent architectural constraints, not a phase — see the security model:
- Never server-execute your graphs.
- Never store or transmit your provider API keys.
- Codegen is one-directional: spec → Python string → download.
Ready? Head to the Quickstart.