CostSentinel

AI cost governance middleware — budget enforcement, attribution, and reporting for LLM API calls.

CostSentinel sits between your application and LLM providers, tracking every token spent, enforcing budget policies, and attributing costs to teams, users, and endpoints. Zero external dependencies beyond PyYAML.

Features

• Token Cost Tracking — Automatic cost calculation for Claude, Titan, and custom models
• Budget Enforcement — Daily/monthly limits with configurable actions (block, downgrade, alert)
• Cost Attribution — Track spending by team, user, endpoint, and model
• Middleware Pattern — Decorator-based interception or manual tracking
• Reporting — CLI and programmatic cost breakdowns
• Zero Infrastructure — JSON file storage for development; DynamoDB-ready for production

Installation

pip install substrai-costsentinel

For AWS integration (DynamoDB state backend):

pip install substrai-costsentinel[aws]

For development:

pip install substrai-costsentinel[dev]

Quickstart

1. Initialize Configuration

costsentinel init

This creates costsentinel.yaml with default policies:

project_name: my-project

pricing:
  claude-3.5-sonnet:
    input: 0.003
    output: 0.015
  claude-3-haiku:
    input: 0.00025
    output: 0.00125

policies:
  global:
    limit_daily: 100.0
    limit_monthly: 2000.0
    on_exceed: block
  team:
    limit_daily: 25.0
    on_exceed: downgrade
  user:
    limit_daily: 5.0
    on_exceed: block
    max_cost_per_request: 0.50

2. Add Middleware to Your Code

from costsentinel import CostMiddleware

middleware = CostMiddleware("costsentinel.yaml")

@middleware.intercept(model="claude-3-haiku", user_id="user-1", team_id="engineering")
def call_llm(prompt):
    # Your LLM call here
    response = my_llm_client.complete(prompt)
    return response.text, response.input_tokens, response.output_tokens

# Returns CallResult with cost info
result = call_llm("Summarize this document...")
print(f"Cost: ${result.cost:.4f}, Remaining: ${result.budget_remaining:.2f}")

3. Manual Tracking

from costsentinel import CostMiddleware

middleware = CostMiddleware("costsentinel.yaml")

# After an LLM call completes
result = middleware.track_call(
    model="claude-3.5-sonnet",
    input_tokens=1500,
    output_tokens=800,
    metadata={"user_id": "user-1", "team_id": "engineering", "endpoint": "/api/chat"}
)

4. Check Reports

costsentinel report --today
costsentinel budget status

Architecture

┌─────────────────────────────────────────────────┐
│                  Your Application                │
├─────────────────────────────────────────────────┤
│              CostSentinel Middleware             │
│  ┌──────────┐  ┌──────────┐  ┌──────────────┐  │
│  │ Pricing  │  │  Budget  │  │ Attribution  │  │
│  │ Engine   │  │ Enforcer │  │    Store     │  │
│  └──────────┘  └──────────┘  └──────────────┘  │
│  ┌──────────────────────────────────────────┐   │
│  │            State Management              │   │
│  │     (JSON file / DynamoDB backend)       │   │
│  └──────────────────────────────────────────┘   │
├─────────────────────────────────────────────────┤
│              LLM Provider (Bedrock)              │
└─────────────────────────────────────────────────┘

CLI Commands

Budget Policies

Policies are evaluated from most specific to least specific:

1. User — Per-user daily/monthly limits
2. Endpoint — Per-API-endpoint limits
3. Team — Per-team limits
4. Global — Organization-wide limits

Actions when budget is exceeded:

• block — Reject the request with BudgetExceededError
• downgrade — Signal to use a cheaper model
• alert — Allow but emit a warning

Development

git clone https://github.com/substrai/costsentinel.git
cd costsentinel
python -m venv .venv
source .venv/bin/activate
pip install -e ".[dev]"
pytest

License

MIT — Copyright (c) 2024 Gaurav Kumar Sinha