agent-budget

One-line spending control for AI agents.

@budget(max_usd=10.00)
async def my_agent(query: str):
    ...  # anything inside stays under $10

Tracks LLM API costs (Anthropic, OpenAI, Gemini, and more) plus x402 payments.
Raises BudgetExceeded the moment your agent tries to go over.

---

Install

pip install bonanza-labs-agent-budget
# PyPI name (CLI/import unchanged). From git:
# pip install "git+https://github.com/c6zks4gssn-droid/agent-budget.git"

No required dependencies. Works with any LLM SDK.

---

Quickstart

from agent_budget import budget, BudgetExceeded
from agent_budget.decorators import track_llm, track_payment, current_spend

@budget(max_usd=5.00)
async def research_agent(query: str) -> str:
    # Call your LLM
    response = await llm.chat(query)

    # Manually record the cost (or use auto_patch=True)
    track_llm("claude-sonnet-4-6",
              input_tokens=response.usage.input_tokens,
              output_tokens=response.usage.output_tokens)

    # Fetch paid content via x402
    data = await fetch_x402("https://data.example.com/report")
    track_payment(0.10, label="x402 report fetch")

    print(f"Spent so far: ${current_spend():.4f}")
    return response.text


# Usage
try:
    result = await research_agent("latest AI research")
except BudgetExceeded as e:
    print(f"Agent hit the limit: spent ${e.spent:.4f} of ${e.limit:.2f}")

---

Auto-patch (zero-instrumentation mode)

Let agent-budget automatically intercept every Anthropic or OpenAI call:

@budget(max_usd=10.00, auto_patch=True)
async def my_agent(query: str):
    # Every client.messages.create() call is tracked — no manual tracking needed
    response = await anthropic_client.messages.create(
        model="claude-sonnet-4-6",
        max_tokens=1024,
        messages=[{"role": "user", "content": query}]
    )
    return response.content[0].text

Install with the extras:

pip install "bonanza-labs-agent-budget[anthropic]"   # auto-patch Anthropic SDK
pip install "bonanza-labs-agent-budget[openai]"      # auto-patch OpenAI SDK
pip install "bonanza-labs-agent-budget[all]"         # both

---

Behaviour options

@budget(
    max_usd=10.00,          # Hard limit in USD
    on_exceed="raise",      # "raise" | "warn" | "ignore"
    auto_patch=True,        # Auto-track Anthropic + OpenAI SDK calls
    report_on_exit=True,    # Print spend report when function exits
)
async def my_agent(): ...

on_exceed	Behaviour
"raise"	Raises BudgetExceeded immediately (default)
"warn"	Emits a RuntimeWarning, continues
"ignore"	Silent — useful for logging/testing

---

Manual tracking (no decorator)

from agent_budget import BudgetTracker

tracker = BudgetTracker(max_usd=5.00)

# Record any LLM call
tracker.track_llm("gpt-4o", input_tokens=1200, output_tokens=400)

# Record an x402 payment
tracker.track_payment(amount_usd=0.25, label="data-purchase")

# Any custom cost
tracker.track_custom(amount_usd=0.01, label="tool-call-fee")

print(tracker.report())
# {
#   "limit_usd": 5.0,
#   "spent_usd": 0.030725,
#   "remaining_usd": 4.969275,
#   "entries": [...]
# }

---

withBudget (functional style)

from agent_budget import withBudget

safe_agent = withBudget(my_agent, max_usd=5.00)
result = await safe_agent("hello")

---

Supported models (built-in price table)

Provider	Models
Anthropic	claude-opus-4-8, claude-sonnet-4-6, claude-haiku-4-5, claude-3.5-*
OpenAI	gpt-4o, gpt-4o-mini, o3, o3-mini, o4-mini, gpt-4-turbo
Google	gemini-2.0-flash, gemini-2.5-pro, gemini-2.5-flash
Meta	llama-3.3-70b, llama-3.1-405b
Mistral	mistral-large, mistral-small

Unknown models fall back to $5.00 / $15.00 per M tokens. Override any price:

from agent_budget import PRICE_TABLE
PRICE_TABLE["my-custom-model"] = {"input": 0.50, "output": 1.50}

---

x402 integration

agent-budget tracks x402 payments when you call track_payment().
For full spending firewall enforcement on x402 payments, see
bonanza-labs.com — production-grade policy engine
with human approval queues, vendor blocklists, and audit logs.

---

FAQ

Does this make network calls?
No. The price table is entirely inline. Zero network calls, zero telemetry.

Does this work with streaming?
Yes for manual tracking — call track_llm() with the final token counts from the stream's usage object. Auto-patch doesn't cover streaming yet (coming in 0.2.0).

What if I nest @budget decorators?
Each call to a @budget-decorated function gets its own isolated context.
Inner contexts don't affect outer ones — you can nest safely.

Thread-safe?
Yes. Uses contextvars.ContextVar for async isolation and threading.Lock inside each context.

---

License

Apache 2.0 — see LICENSE.

Built by Bonanza Labs.
For teams: managed dashboard, policy engine, and approval queue at bonanza-labs.com/firewall.