AgentWallet

Financial governance infrastructure for AI agents.

Spend controls, rules engine, kill switch, and audit trail — every financial action an agent takes flows through governance before a dollar moves.

Install

pip install agentwallet

Quick Start

from agentwallet import AgentWallet, SpendRule, RuleVerdict

# Create a governed wallet
wallet = AgentWallet("trading-agent", budget_cents=5000)  # $50 budget

# Add custom governance rules
wallet.add_rule(SpendRule(
    rule_id="block-expensive-models",
    name="Block calls over $5",
    condition=lambda ctx: ctx["amount_cents"] > 500
        and "gpt-4" in ctx.get("metadata", {}).get("model", ""),
    verdict=RuleVerdict.DENY,
))

# Every spend goes through governance
result = wallet.spend(200, "llm-inference", metadata={"model": "gpt-4"})
# {'approved': True, 'tx_id': '...', 'remaining_cents': 4800}

result = wallet.spend(800, "llm-inference", metadata={"model": "gpt-4"})
# {'approved': False, 'reason': 'Blocked by rule: block-expensive-models'}

Persistence (NEW in 0.2.0)

Add persist=True and wallet state survives process restarts. Zero config — uses SQLite under the hood.

# First run
wallet = AgentWallet("my-agent", budget_cents=5000, persist=True)
wallet.spend(200, "api-call")
wallet.spend(300, "web-search")
print(wallet.balance_cents)  # 4500

# ... restart your process ...

# Second run — state is restored automatically
wallet = AgentWallet("my-agent", persist=True)
print(wallet.balance_cents)  # 4500 — it remembered!
print(len(wallet.transactions))  # 2 — transactions restored too

Everything persists: balance, transactions, audit log, kill switch state. Multiple agents can share the same database file.

# Custom database path
wallet = AgentWallet("agent-1", persist=True, db_path="/data/governance.db")

# Query spend analytics (with persistence)
summary = wallet._storage.get_spend_summary("agent-1")
by_category = wallet._storage.get_spend_by_category("agent-1")
all_wallets = wallet._storage.list_wallets()

What's Included

• AgentWallet — Per-agent wallets with configurable budgets
• GovernanceEngine — Rules engine that evaluates every transaction
• SpendRule — Custom rules with priority-based evaluation
• Kill Switch — Instant shutdown of all agent spending
• AuditLog — Append-only JSONL audit trail for compliance
• SQLite Persistence — persist=True and state survives restarts
• Spend Analytics — Query spend by category, get summaries
• Event Callbacks — wallet.on("deny", callback) for webhooks/alerts
• Dashboard API — FastAPI server for real-time monitoring
• CLI — agentwallet demo and agentwallet dashboard

Built-in Safety Rules

Every wallet ships with four default rules (highest priority first):

1. Kill Switch (priority 1000) — blocks everything when activated
2. Max Per Transaction (priority 900) — caps individual spend
3. Daily Limit (priority 800) — caps rolling 24-hour spend
4. Balance Check (priority 700) — prevents overdraft

Event Callbacks

Get notified when governance events fire:

import requests

# Webhook on denied transactions
wallet.on("deny", lambda data: requests.post(
    "https://hooks.slack.com/your-webhook",
    json={"text": f"🚫 Agent blocked: {data['reason']}"}
))

# Log all approved spend
wallet.on("approve", lambda data: print(f"✅ ${data['amount_cents']/100:.2f} approved"))

# Kill switch alerts
wallet.on("kill_switch", lambda data: print(f"🛑 Kill switch: {data}"))

Dashboard

pip install agentwallet[dashboard]
agentwallet dashboard

Or programmatically:

from agentwallet import AgentWallet, register_wallet, start_dashboard_server

wallet = AgentWallet("my-agent", budget_cents=10000)
register_wallet(wallet)
start_dashboard_server(port=8100)  # API at http://localhost:8100/docs

CLI

agentwallet demo        # Run a quick demo with sample transactions
agentwallet dashboard   # Start the governance API server
agentwallet version     # Show version

Zero Dependencies

The core SDK has zero external dependencies. Just Python 3.9+. The dashboard server optionally uses FastAPI + uvicorn.

Links

• Quickstart Repo: github.com/JackD720/agentwallet-quickstart
• Examples: 3 working examples (simple agent, LangChain, CrewAI)
• Reference: arXiv:2501.10114 "Infrastructure for AI Agents"

License

MIT