AI agent integration

Build agents that move money safely

Alloy is AI-native: every API is designed for autonomous agent consumption with self-discoverable endpoints, structured errors, scoped authority, and human escalation built in.

Design principles

Self-discovery

Module schemas with x-agent-hints annotations. Agents discover endpoints, parameters, and constraints without documentation lookup.

Same controls

Agents go through the same policy, risk, and compliance checks as human operators. No shortcuts, no backdoors.

Scoped authority

Delegated tokens with explicit scope, amount caps, asset restrictions, and time-bounded expiry. Agents get the minimum authority they need.

Human escalation

High-risk or out-of-policy actions automatically escalate to human approvers. Agents propose; humans approve when policy demands it.

MCP server integration

Alloy exposes a Model Context Protocol server for Claude, GPT, and other MCP-compatible agents.

Configuration

{
  "mcpServers": {
    "alloy": {
      "url": "https://api.alloy.build/mcp/v1",
      "transport": "streamable-http",
      "headers": {
        "Authorization": "Bearer ${ALLOY_AGENT_TOKEN}"
      }
    }
  }
}

Available MCP tools

list_wallets List wallets with balances across providers

get_wallet Get detailed wallet information

create_transaction Submit a policy-bound transaction intent

check_policy Pre-check a transaction against policies

get_risk_score Get risk assessment for an address or amount

list_events Query recent events and alerts

get_recon_status Check reconciliation status

Delegated authority

Create scoped agent tokens with explicit permissions, amount caps, and asset restrictions. Every action is logged with the agent identity for full auditability.

Create an agent token

curl -X POST https://api.alloy.build/v1/agent-tokens \
  -H "Authorization: Bearer $ALLOY_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "name": "payment-agent",
    "scopes": [
      "wallets:read",
      "transactions:create",
      "transactions:read"
    ],
    "constraints": {
      "max_amount_per_tx": "5000.00",
      "allowed_assets": ["USDC", "USDT"],
      "require_policy_check": true
    },
    "expires_in": "24h"
  }'

Available scopes

Scope	Description	Risk
wallets:read	List wallets, get balances	Low
wallets:create	Create new wallets	Medium
transactions:read	View transaction history	Low
transactions:create	Submit transaction intents	High
policies:read	View policy configurations	Low
policies:evaluate	Trigger policy evaluation	Medium
risk:read	View risk scores and alerts	Low
recon:read	View reconciliation data	Low
events:subscribe	Subscribe to event streams	Low

Agent patterns

Payment agent

An autonomous agent that processes vendor payments with policy-bound transaction intents and automatic human escalation.

from alloy import AlloyClient

client = AlloyClient(
    api_key=os.environ["ALLOY_AGENT_TOKEN"],
    base_url="https://api.alloy.build"
)

# Agent discovers available wallets
wallets = client.wallets.list()

# Submit a policy-bound transaction intent
intent = client.transaction_intents.create(
    wallet_id=wallets[0].id,
    asset="USDC",
    amount="1000.00",
    destination="0x742d35Cc6634C0532925a3b844Bc9e7595f2bD18",
    policy="agent-low-risk-transfer",
    metadata={
        "agent_id": "payment-bot-v2",
        "reason": "Vendor invoice #INV-2024-0847"
    }
)

# Handle policy escalation
if intent.status == "pending_human_approval":
    print(f"Escalated to human: {intent.escalation_reason}")
else:
    print(f"Approved: {intent.id} -> {intent.status}")

Monitoring agent

A read-only agent that streams events, detects anomalies, and alerts the team — without any write permissions.

import alloy
import asyncio

client = alloy.AlloyClient(api_key=os.environ["ALLOY_AGENT_TOKEN"])

async def monitor():
    async for event in client.events.stream(
        types=["transaction.completed", "transaction.failed",
               "risk.alert", "policy.violation"]
    ):
        if event.type == "risk.alert":
            # Agent can read risk details but cannot override
            alert = client.risk.get_alert(event.data.alert_id)
            await notify_team(alert)

        elif event.type == "policy.violation":
            # Log violation evidence for audit
            violation = client.policies.get_decision(
                event.data.decision_id
            )
            await log_compliance_event(violation)

asyncio.run(monitor())

Security model

Layer	Mechanism
Authentication	Scoped bearer tokens with expiry, IP allowlist, and revocation
Authorization	Fine-grained scopes per resource type; deny-by-default
Policy enforcement	Every write action passes through PolicyKit evaluation
Amount limits	Per-transaction and rolling-window caps enforced at token level
Human escalation	Out-of-policy actions pause and notify human approvers
Audit trail	Every agent action logged with agent identity, scope, and policy receipt

Framework compatibility

Claude MCP

Native MCP server with streamable HTTP transport.

LangChain / LangGraph

Schema-based tool definition with structured output.

CrewAI

Agent tool integration via module schema.

OpenAI Assistants

Function calling with JSON Schema definitions.

Vercel AI SDK

TypeScript tool definitions with streaming support.

Coinbase AgentKit

Wallet operations bridge for onchain agents.