🛡️ TokenShield

Real-time token cost monitoring, budget enforcement, and optimization for LLM applications.

Stop burning money on LLM API calls. TokenShield gives you per-request cost tracking, budget gates, and automatic optimization — before the invoice arrives.

The Problem

Month 1:  $50    "This is cheap!"
Month 2:  $200   "Growth is normal"
Month 3:  $3,400 "WHAT HAPPENED?!"

LLM costs are invisible until the bill arrives. A single misconfigured loop, a verbose system prompt, or an unbound tool list can 10x your spend overnight.

The Solution

from tokenshield import Shield, BudgetPolicy

shield = Shield(
    model="gpt-4o",
    policy=BudgetPolicy(
        max_cost_per_request=0.05,     # $0.05 per request
        max_cost_per_hour=2.00,        # $2/hour
        max_cost_per_day=20.00,        # $20/day
        alert_threshold_pct=80,        # Alert at 80% of any limit
    )
)

# Wrap any LLM call
result = shield.call(
    messages=[{"role": "user", "content": "Summarize this order"}],
    tools=tool_schemas,
)

print(shield.report())
# ┌─────────────────────────────────┐
# │ Requests today:     142         │
# │ Tokens (in/out):    89K / 12K   │
# │ Cost today:         $4.23       │
# │ Budget remaining:   $15.77      │
# │ Avg cost/request:   $0.030      │
# │ Most expensive:     search (48%)│
# └─────────────────────────────────┘

Features

Architecture

┌──────────────────────────────────────────────────────────┐
│                     Your Application                      │
├──────────────────────────────────────────────────────────┤
│                                                          │
│  ┌──────────┐   ┌──────────┐   ┌──────────────────────┐ │
│  │ shield   │──→│ estimator│──→│ budget_gate          │ │
│  │ .call()  │   │ (tokens) │   │ (allow / reject)     │ │
│  └──────────┘   └──────────┘   └──────────┬───────────┘ │
│       │                                     │            │
│       │         ┌──────────┐   ┌───────────▼──────────┐ │
│       │         │ tracker  │←──│ LLM API call         │ │
│       │         │ (costs)  │   │ (litellm / openai)   │ │
│       │         └────┬─────┘   └──────────────────────┘ │
│       │              │                                   │
│  ┌────▼──────────────▼─────┐   ┌──────────────────────┐ │
│  │ reporter                │   │ alert_hooks          │ │
│  │ (dashboard / export)    │   │ (webhook / callback) │ │
│  └─────────────────────────┘   └──────────────────────┘ │
│                                                          │
└──────────────────────────────────────────────────────────┘

Quick Start

pip install tokenshield

Basic Usage

from tokenshield import Shield

shield = Shield(model="gpt-4o")

# Track a call (wrap your existing LLM call)
result = shield.call(messages=[...])

# Check current spend
print(f"Today: ${shield.tracker.cost_today:.2f}")

Budget Enforcement

from tokenshield import Shield, BudgetPolicy

shield = Shield(
    model="gpt-4o",
    policy=BudgetPolicy(max_cost_per_request=0.10)
)

try:
    result = shield.call(messages=huge_prompt)
except shield.BudgetExceeded as e:
    print(f"Blocked! Estimated cost ${e.estimated_cost:.3f} exceeds limit")

Alert Hooks

shield = Shield(
    model="gpt-4o",
    policy=BudgetPolicy(max_cost_per_day=20.00, alert_threshold_pct=80),
    on_alert=lambda msg: slack.post(channel="#llm-costs", text=msg),
)

Optimization Suggestions

tips = shield.optimize(messages, tools)
# [
#   "System prompt is 3,800 tokens (63% of input). Consider compressing.",
#   "18 tools bound but only 3 used. Use dynamic tool binding to save ~2,250 tokens.",
#   "History has 45 messages. Consider windowing to last 20.",
# ]

Pricing Database

Built-in pricing (updated monthly):

Add custom models:

shield.pricing.add("my-finetuned-model", input=5.00, output=15.00)

Documentation

• Architecture & Data Flow — Mermaid diagrams of the full pipeline
• Benchmarks — Cost savings measurements across real workloads
• API Reference — Full class/method documentation

License

MIT — see LICENSE