nullspend 0.4.0

Python SDK for NullSpend — FinOps for AI agents

nullspend

Python SDK for NullSpend — FinOps for AI agents.

Installation

pip install nullspend

Quick Start

from nullspend import NullSpend, CostEventInput

ns = NullSpend(
    base_url="https://nullspend.dev",
    api_key="ns_live_sk_...",
)

# Report a cost event
ns.report_cost(CostEventInput(
    provider="openai",
    model="gpt-4o",
    input_tokens=1200,
    output_tokens=350,
    cost_microdollars=5250,
    tags={"environment": "production", "agent": "support-bot"},
))

# Check budget status
status = ns.check_budget()
for entity in status.entities:
    print(f"{entity.entity_type}/{entity.entity_id}: "
          f"${entity.remaining_microdollars / 1_000_000:.2f} remaining")

Features

• Cost event reporting (single and batch)
• @track_tool decorator + track() inline for tool cost tracking
• Budget status and listing
• Cost analytics summaries
• Human-in-the-loop action management (create, poll, mark result)
• propose_and_wait() high-level orchestrator
• Tracked httpx transport with automatic cost tracking and enforcement
• Loop detection for stuck agents (client-side and proxy-side)
• Automatic retries with exponential backoff
• Idempotency keys on mutating requests
• Type hints throughout (py.typed)

Tool Cost Tracking

Track costs for non-LLM tools (API calls, web searches, database queries) with a decorator or inline:

from nullspend import NullSpend

ns = NullSpend(api_key="ns_live_sk_...", base_url="https://nullspend.dev")

# Decorator — reports cost after function executes, measures duration
@ns.track_tool(cost=0.02, tool_name="web_search")
def search(query: str) -> str:
    return requests.get(f"https://api.example.com/search?q={query}").text

# Inline — reports cost and returns the result
result = ns.track(call_api(), cost=0.01, tool_name="search")

Costs are reported even if the tool raises an exception (the API call happened, tokens were consumed). Failed calls are tagged with _ns_error: "true" for dashboard filtering.

Optional fields: provider, model, tool_server, tags, customer. Uses batch reporter when configured, falls back to direct reporting.

Loop Detection

Detects agents stuck in infinite loops — repeated identical calls that burn budget without progress.

Proxy users: Loop detection is on by default. If your agent calls the same model with identical content 50+ times in 60 seconds, the proxy returns a 429 with code: "loop_detected". No configuration needed.

SDK users: Opt in with one line:

from nullspend import create_tracked_client

client = create_tracked_client("openai", loop_detection=True)

Customize thresholds:

from nullspend import create_tracked_client, LoopDetectionConfig

client = create_tracked_client("openai", loop_detection=LoopDetectionConfig(
    max_calls=100,       # higher for batch workloads
    window_seconds=120,  # wider window
))

Disabling: Set loop_max_calls=0 on the budget entity via the API or dashboard.

Error handling:

Error	Code	When
LoopDetectedError	loop_detected	Same model+content called 50+ times in 60s
BudgetExceededError	budget_exceeded	Budget exhausted
VelocityExceededError	velocity_exceeded	Spend rate exceeds velocity limit
SessionLimitExceededError	session_limit_exceeded	Session spend cap reached
TagBudgetExceededError	tag_budget_exceeded	Tag-level budget exhausted
MandateViolationError	mandate_violation	Model/provider not allowed

from nullspend import LoopDetectedError, BudgetExceededError

try:
    response = client.chat.completions.create(...)
except LoopDetectedError as e:
    print(f"Loop detected: {e.model} called {e.call_count} times")
    print(f"Detection type: {e.detection_type}")  # "per_key" or "aggregate"
except BudgetExceededError as e:
    print(f"Budget exceeded: {e.remaining_microdollars} microdollars remaining")
    if e.recovery:
        print(f"Retryable: {e.recovery['retryable']}")
        print(f"Owner action required: {e.recovery['owner_action_required']}")

Every denial error includes an optional recovery dict with machine-readable hints:

Field	Type	Meaning
retryable	bool	Whether the request can succeed if retried later
owner_action_required	bool	Whether a human or config change is needed
retry_after_seconds	int | None	Seconds to wait before retry (retryable denials only)
docs	str | None	Documentation URL for this error type

recovery is None when connecting to an older proxy that doesn't include it.

Documentation

See the NullSpend docs for full API reference.