Policy Engine

Overview

The Sardis Policy Engine is the core enforcement layer that sits between payment requests and chain execution. Every transaction must pass policy checks before it can be signed and broadcast.

Payment Request → Policy Engine → MPC Signing → Chain Execution

The policy engine evaluates:

• Spending limits (per-transaction, daily, weekly, monthly)
• Merchant rules (allowlists, blocklists, category restrictions)
• Amount thresholds (auto-approve vs human review)
• Time-based rules (business hours only, no weekends)
• Compliance checks (sanctions screening, risk scoring)

Natural Language Policies

Define policies in plain English. The policy engine parses and enforces them automatically:

from sardis import SardisClient

client = SardisClient(api_key="sk_...")

wallet = client.wallets.create(
    name="research-agent",
    chain="base",
    token="USDC",
    policy="Max $100/day, Max $25 per transaction, only OpenAI and Anthropic"
)

Supported Policy Expressions

Structured Policies

For programmatic control, use structured JSON policies:

policy = {
    "limits": {
        "per_transaction": 5000,
        "daily": 20000,
        "monthly": 100000,
    },
    "merchants": {
        "allowed": ["aws.amazon.com", "openai.com"],
        "blocked_categories": ["gambling", "adult"],
    },
    "approval": {
        "auto_below": 1000,
        "human_above": 1000,
    },
    "time": {
        "allowed_hours": {"start": 9, "end": 17},
        "timezone": "America/New_York",
        "block_weekends": True,
    },
}

wallet = client.wallets.create(
    name="ops-agent",
    chain="base",
    token="USDC",
    policy=policy,
)

Policy Evaluation

When a payment is requested, the policy engine runs through this evaluation pipeline:

1. Parse — Natural language or JSON policy is parsed into rules
2. Check limits — Per-tx, daily, weekly, monthly against current usage
3. Check merchants — Destination against allowlist/blocklist
4. Check time — Current time against time restrictions
5. Check compliance — Sanctions screening (Elliptic), risk scoring
6. Determine approval — Auto-approve or require human review
7. Return verdict — Allow, deny, or require approval with reason

# Check a transaction against policy before executing
verdict = wallet.check_policy(
    amount=250,
    destination="openai.com",
    purpose="GPT-4 API credits",
)

print(verdict.allowed)       # True
print(verdict.reason)        # "Within all policy limits"
print(verdict.requires_human) # False

MCP Policy Tools

The MCP server exposes policy tools for Claude and Cursor:

Fail-Closed Design

The policy engine defaults to deny on any failure:

• If the policy cannot be parsed, the transaction is denied
• If compliance screening times out, the transaction is denied
• If the merchant cannot be verified, the transaction is denied
• If any rule evaluation fails, the transaction is denied

This fail-closed approach ensures that bugs or outages never result in unauthorized payments.