letagentpay 0.2.3

Python SDK for LetAgentPay — AI agent spending policy middleware

letagentpay

Python SDK for LetAgentPay — AI agent spending policy middleware. Set budgets, define spending policies, and control AI agent purchases.

Installation

pip install letagentpay

Quick Start

from letagentpay import LetAgentPay

client = LetAgentPay(token="agt_xxx")

# Create a purchase request
result = client.request_purchase(
    amount=15.0,
    category="api_calls",
    description="OpenAI GPT-4 call",
)
print(result.status)  # "auto_approved" / "pending" / "rejected"

# Check budget
budget = client.check_budget()
print(f"Remaining: ${budget.remaining}")

@guard Decorator

Automatically check the spending policy before executing a function:

from letagentpay import guard

@guard(token="agt_xxx", category="api_calls", amount=0.03)
def call_openai(prompt: str) -> str:
    return openai.chat.completions.create(
        model="gpt-4",
        messages=[{"role": "user", "content": prompt}]
    ).choices[0].message.content

# Runs only if the policy allows it
result = call_openai("Analyze this document")

Environment Variables

export LETAGENTPAY_TOKEN=agt_xxx
export LETAGENTPAY_BASE_URL=https://api.letagentpay.com/api/v1/agent-api  # optional

# Token is taken from LETAGENTPAY_TOKEN
client = LetAgentPay()

Security Model

LetAgentPay uses server-side cooperative enforcement. When your agent calls request_purchase(), the request is evaluated by the policy engine on the LetAgentPay server. The agent receives only the result (approved/denied/pending) and cannot:

• Modify its own policies (the agt_ token grants access only to the Agent API)
• Override policy check results (they come from the server)
• Approve its own pending requests (only a human can do that via the dashboard)

This is a cooperative model — it protects against budget overruns, category violations, and scheduling mistakes by well-behaved agents. It does not sandbox a malicious agent that has direct access to payment APIs.

Best Practices

• Don't give your agent raw payment credentials (Stripe keys, credit card numbers). LetAgentPay should be the only spending channel
• Use pending + manual approval for high-value purchases
• Set per-request limits as an additional barrier
• Review the audit trail in the dashboard regularly

Documentation

• Python SDK docs
• Agent API Reference
• GitHub

License

MIT