spendpol 0.2.0

Spendpol — AI Spend Protocol. Delegated spending governance for autonomous AI agents.

🔐 Spendpol

AI Spend Protocol — Governance for autonomous agent spending.

Spendpol is an open protocol that sits between AI agents and payment systems. It answers one question:

"Should this agent be allowed to make this expenditure?"

Spendpol does not move money. It governs whether money should move.

---

Why Spendpol?

AI agents are calling APIs, purchasing compute, and consuming paid services. But spending controls are stuck in the human era: hardcoded API keys, manual approval, cloud budget alerts.

As agents scale from 1 to 1,000 and delegate tasks to sub-agents, the uncontrolled spending surface grows exponentially. Spendpol provides the governance layer.

Agent Logic → [Spendpol] → Payment Execution
                 │
                 ├── Is this agent authorized?
                 ├── Within spending limits?
                 ├── Anomaly detected?
                 └── Audit logged ✓

Install

pip install spendpol

Quick Start

from spendpol import SpendpolClient

client = SpendpolClient()

# Human grants spending authority to an agent
client.grant(
    delegator="user:[e-mail]",
    delegatee="agent-research",
    max_per_transaction=1.00,
    max_daily=10.00,
    max_monthly=200.00,
    allowed_vendors=["openai", "anthropic"],
    allowed_actions=["api_call"],
)

# Agent requests permission to spend
verdict = client.check_intent(
    agent_id="agent-research",
    vendor="openai",
    action="api_call",
    amount=0.03,
    purpose="Generate embeddings for ticket #4521",
)

if verdict.allowed:
    result = call_openai_api()  # Your payment execution
    client.log_receipt(verdict, actual_amount=0.028)
elif verdict.denied:
    print(f"Blocked: {verdict.evaluations[-1].detail}")

CrewAI Integration

from spendpol.crewai import SpendpolCrewMiddleware

asp = SpendpolCrewMiddleware(
    delegator="user:[e-mail]",
    default_daily_limit=10.00,
    allowed_vendors=["openai", "anthropic"],
)

# Decorator: one line to control any tool's spending
@asp.controlled(vendor="openai", action="api_call", cost_per_call=0.03)
def embedding_tool(text: str) -> list[float]:
    return openai.embeddings.create(input=text)

# Automatically: checks policy → executes → logs receipt
result = embedding_tool("Hello world")

Core Concepts

Concept	What it does
SpendIntent	Agent declares "I want to spend $X on Y"
PolicyVerdict	Engine responds: allow / deny / escalate
SpendReceipt	Records what actually happened
AuditQuery	Anyone can inspect spending history
DelegationGrant	Human gives agent permission with limits
Attenuation	Sub-agents can only have fewer permissions, never more

What Spendpol Is NOT

• Not a wallet — holds no funds
• Not a payment processor — moves no money
• Not a blockchain — no consensus, no tokens
• Not KYC/AML — no identity verification
• Not a legal entity — agents have no personhood

Spendpol's regulatory surface is equivalent to a logging tool, not a financial service.

Architecture

spendpol/
├── schemas.py             # Protocol message types
├── delegation.py          # Delegation chains + attenuation
├── wallet.py              # Non-custodial spending tracker
├── policy.py              # Deterministic policy engine
├── audit.py               # Append-only audit log
├── validator.py           # Message validation + verification
├── service.py             # Agent-to-agent commerce registry
├── client.py              # High-level SDK
├── crewai.py              # CrewAI middleware
├── langgraph.py           # LangGraph middleware
├── autogen.py             # AutoGen middleware
├── schemas/               # JSON Schema definitions (Draft 2020-12)
│   ├── spend-intent.schema.json
│   ├── policy-verdict.schema.json
│   ├── spend-receipt.schema.json
│   ├── delegation-grant.schema.json
│   ├── audit-query.schema.json
│   ├── cost-estimate.schema.json
│   ├── service-offer.schema.json
│   ├── service-quote.schema.json
│   └── service-result.schema.json
├── examples/              # Test vector payloads
│   ├── spend-intent.json
│   ├── policy-verdict-allow.json
│   ├── policy-verdict-deny.json
│   ├── spend-receipt.json
│   ├── delegation-grant.json
│   ├── audit-query.json
│   ├── service-offer.json
│   ├── service-quote.json
│   └── service-result.json
├── test_core.py           # 67 core tests
├── test_langgraph.py      # 22 LangGraph tests
├── test_autogen.py        # 23 AutoGen tests
└── WHITEPAPER.md          # Full protocol specification

Run Tests

pip install pytest
python -m pytest test_core.py test_langgraph.py test_autogen.py -v

Protocol Specification

Read the full WHITEPAPER.md for:

• Complete message schemas with JSON examples
• Delegation model and attenuation principle
• Policy engine evaluation order
• Anomaly detection specification
• Trust and liability model
• Audit role taxonomy

Roadmap

• Core protocol (4 message types)
• Delegation chains with attenuation
• Deterministic policy engine
• Anomaly detection (confidence-aware)
• Audit log with role-based access
• CrewAI integration
• LangGraph integration
• AutoGen integration
• REST API server (spendpol-cloud)
• Dashboard (spend analytics)
• PyPI publish
• Agent-to-agent commerce (Phase 1)

Contributing

Spendpol is an open protocol. Contributions welcome.

1. Fork the repo
2. Create your feature branch (git checkout -b feature/my-feature)
3. Run tests (python -m pytest test_core.py test_langgraph.py test_autogen.py -v)
4. Submit a PR

License

• Python code (reference implementation): MIT
• Protocol specification (WHITEPAPER.md, JSON Schemas): CC-BY-4.0

See DISCLAIMER.md for important legal notices.

---

Cloud computing gave machines compute. Spendpol gives machines governance.