Skip to main content

BudgetGate

Deterministic, pre-execution spend limiting for semantic actions in agent systems.

Source of Truth

The canonical source is github.com/actiongate-oss/budgetgate. PyPI distribution is a convenience mirror.

Vendoring encouraged. This is a small, stable primitive. Copy it, fork it, reimplement it. See SEMANTICS.md for the behavioral contract if you reimplement.


Quick Start

from decimal import Decimal
from budgetgate import Engine, Ledger, Budget, BudgetExceeded

engine = Engine()

@engine.guard(
    Ledger("openai", "gpt-4", "user:123"),
    Budget(max_spend=Decimal("10.00"), window=3600),  # $10/hour
    cost=Decimal("0.03"),  # fixed cost per call
)
def call_gpt4(prompt: str) -> str:
    return openai.chat(prompt)

try:
    response = call_gpt4("Hello")
except BudgetExceeded as e:
    print(f"Budget exceeded: {e.decision.spent_in_window} spent")

Two Cost Modes

Fixed Cost (pre-execution)

When cost is known before execution:

@engine.guard(
    Ledger("openai", "embedding"),
    Budget(max_spend=Decimal("5.00"), window=3600),
    cost=Decimal("0.0001"),  # fixed cost per call
)
def embed(text: str) -> list[float]:
    return openai.embed(text)

Bounded Dynamic Cost (pre-execution with estimate)

When cost depends on the result but has a known upper bound:

@engine.guard_bounded(
    Ledger("anthropic", "claude", "user:123"),
    Budget(max_spend=Decimal("5.00"), window=3600),
    estimate=Decimal("0.50"),  # max possible cost (reserved before execution)
    actual=lambda r: Decimal(str(r.usage.total_cost)),  # actual cost (committed after)
)
def call_claude(prompt: str) -> Response:
    return anthropic.messages.create(...)

The estimate is reserved before execution. If it doesn't fit the budget, the action is blocked. After execution, the actual cost is committed and unused budget is recovered.


Core Concepts

Ledger

Identifies a spend-tracked stream:

Ledger(namespace, resource, principal)

Ledger("openai", "gpt-4", "user:123")     # per-user
Ledger("anthropic", "claude", "team:eng") # per-team
Ledger("infra", "compute", "global")      # global

Budget

Budget(
    max_spend=Decimal("10.00"),  # max spend in window
    window=3600,                  # rolling window (seconds)
    mode=Mode.HARD,               # HARD raises, SOFT returns result
    on_store_error=StoreErrorMode.FAIL_CLOSED,
)

Decision

Every check returns a Decision with:

decision.allowed          # bool
decision.spent_in_window  # Decimal - current spend
decision.remaining        # Decimal - budget remaining
decision.requested        # Decimal - amount requested

Decorator Styles

Decorator Cost Mode Returns On Block
guard Fixed T Raises BudgetExceeded
guard_bounded Dynamic T Raises BudgetExceeded
guard_result Fixed Result[T] Returns blocked result
guard_bounded_result Dynamic Result[T] Returns blocked result
# Raises on block
@engine.guard(ledger, budget, cost=Decimal("0.01"))
def fixed_action(): ...

@engine.guard_bounded(ledger, budget, estimate=Decimal("0.50"), actual=lambda r: r.cost)
def dynamic_action(): ...

# Never raises - returns Result[T]
@engine.guard_result(ledger, budget, cost=Decimal("0.01"))
def fixed_action(): ...

@engine.guard_bounded_result(ledger, budget, estimate=Decimal("0.50"), actual=lambda r: r.cost)
def dynamic_action(): ...

Relation to ActionGate

BudgetGate complements ActionGate:

Primitive Limits Use case
ActionGate calls/time Rate limiting
BudgetGate cost/time Spend limiting

Both are:

  • Deterministic
  • Pre-execution
  • Decorator-friendly
  • Store-backed

Use together:

from decimal import Decimal

@actiongate_engine.guard(Gate("api", "search"), Policy(max_calls=100))
@budgetgate_engine.guard(Ledger("api", "search"), Budget(max_spend=Decimal("1.00")), cost=Decimal("0.01"))
def search(query: str) -> list:
    ...

API Reference

Type Purpose
Engine Core spend tracking
Ledger Spend stream identity
Budget Spend policy
Decision Evaluation result
Result[T] Wrapper for guard_result
BudgetExceeded Exception from guard
Enum Values
Mode HARD, SOFT
StoreErrorMode FAIL_CLOSED, FAIL_OPEN
Status ALLOW, BLOCK
BlockReason BUDGET_EXCEEDED, STORE_ERROR

Numeric Precision

All spend amounts use Decimal to avoid floating-point drift. See SEMANTICS.md §9.


License

Apache License 2.0. See LICENSE for the full text.

Release files for budgetgate 0.3.1

For a detailed explanation of source distributions (sdists) and built distributions (wheels), please see the package formats documentation.

Source distribution (sdist)

Source distribution for budgetgate 0.3.1
File Size Uploaded
budgetgate-0.3.1.tar.gz 24.3 kB Details

Built distribution (wheel)

Table of built distributions (wheels) for budgetgate 0.3.1
File Interpreter ABI Platform
budgetgate-0.3.1-py3-none-any.whl Python 3 none any Details

Total release size: 41.2 kB

Release files / budgetgate-0.3.1.tar.gz

Download URL budgetgate-0.3.1.tar.gz
Size 24.3 kB
Tags Source
SHA-256 checksum
How to use checksums
c0f2e08450f2ced988eb8ea9c921a785632ea1b5848ebcb224634943c77ac64d
BLAKE2b-256 checksum
How to use checksums
86ca00b2cb03e079210eb7fa1e73dd74eb78f3affea08c93cd4c9e2a40c41df7
Upload date
Uploaded using Trusted Publishing?
What is trusted publishing?
No
Uploaded via twine/6.2.0 CPython/3.14.0

Release files / budgetgate-0.3.1-py3-none-any.whl

Download URL budgetgate-0.3.1-py3-none-any.whl
Size 16.9 kB
Tags Python 3
SHA-256 checksum
How to use checksums
bb004c78184a01427cab682ca6e1b0cec1017fd938b5198b5495776f2c23905d
BLAKE2b-256 checksum
How to use checksums
573112aa07be403e59130134f68236abff4e4ecfb934bda6caec82ec82fd28f5
Upload date
Uploaded using Trusted Publishing?
What is trusted publishing?
No
Uploaded via twine/6.2.0 CPython/3.14.0

Release history Release notifications | RSS feed

This release

0.3.1 This release

2 release files

0.3.0

2 release files

0.2.2

2 release files

0.2.1

2 release files

0.2.0

2 release files

Anthropic, PBC Visionary sponsor Bloomberg Visionary sponsor Hudson River Trading Visionary sponsor Meta Visionary sponsor NVIDIA Visionary sponsor Microsoft Sustainability sponsor Depot Continuous Integration AWS Cloud computing and Security Sponsor Datadog Monitoring Fastly CDN Google Download Analytics Sentry Error logging StatusPage Status page