agentguard-protocol

agentguard-protocol

The governance layer for autonomous AI agent payments.

pip install agentguard-protocol

Every .pay() call runs through five deterministic layers — kill switch, ERC-8004 identity, policy, anomaly, Claude-powered intent — before settling as a USDC Nanopayment on Arc and writing an on-chain audit receipt. Three lines of SDK code replaces a raw Circle Wallets transfer, and the agent can't bypass the governance layer without replacing the keys.

---

Why this exists

In 2026, AI agents will handle money. The current options are both wrong:

• Give the agent raw wallet keys — no limits, no audit, one prompt injection away from a drained treasury.
• Human-in-the-loop every payment — kills the autonomy the agent was built for, doesn't scale to sub-cent operational payments.

AgentGuard sits between the agent and the rail. The agent gets autonomy. The operator gets a policy file, an audit log, and a kill switch.

Three-line integration

from agentguard import AgentGuard

guard = AgentGuard(
    agent_id="research-agent-v1",
    policy_file="policies/research-agent-v1.yaml",
    api_base_url="http://localhost:4000",
)

receipt = guard.pay(
    to_agent_id="data-vendor-agent-v1",
    amount_usdc=0.001,
    intent="Buy Q3 macro stats report",
    original_task_id="brief-2026-w17",
)

if receipt.approved:
    print(f"settled {receipt.amount_usdc} USDC — arc tx: {receipt.arc_tx_hash}")
else:
    print(f"blocked: {receipt.reason}")

What runs on every .pay()

A blocked payment never settles. An approved payment settles and logs a receipt. Either way, the full trace is on-chain.

Install

The base SDK:

pip install agentguard-protocol

With real Circle Gateway + x402 settlement (pulls in circle-titanoboa-sdk):

pip install "agentguard-protocol[circle]"

For development (tests, linting):

pip install "agentguard-protocol[dev]"

Running the governance API

The SDK is a client — it talks to the AgentGuard governance API that runs the actual pipeline. You host the API yourself:

git clone https://github.com/vikramRooT/agentguard
cd agentguard
pnpm install
pnpm infra:up            # Postgres + Redis via docker compose
pnpm dev:api             # Node API on :4000
pnpm seed                # register demo agents

Then point the SDK at http://localhost:4000. Full deployment guide: /docs/self-hosting.md.

API reference

AgentGuard(agent_id, policy_file=..., policy=..., api_base_url=..., api_key=..., auto_register=True, private_key=None, circle_wallet_id=None, circle_wallet_address=None, chain="arcTestnet")

.pay(*, to_agent_id=None, to_wallet_address=None, amount_usdc, intent, original_task_id=None, context=None, strict=False) -> PaymentReceipt

The primary A2A transfer. Exactly one of to_agent_id / to_wallet_address is required. intent is free-form — used by the intent classifier and surfaced on the dashboard.

When strict=True, non-approvals raise PaymentBlocked. Otherwise a PaymentReceipt is returned with approved=False.

.pay_for_service(*, url, intent, expected_max_usdc=10.0, ...) -> PaymentReceipt

Pay an x402-protected HTTP endpoint. Runs the governance pipeline, then — if approved and circlekit is installed — actually settles via Circle Gateway. Receipt's .evidence["settlement"] carries the on-chain proof.

.escalate(receipt)

Flag a blocked receipt for human review via the API.

.balances() -> dict | None

Returns wallet + Gateway balances if a settler is configured, else None.

.close() / context manager

with AgentGuard(...) as guard:
    guard.pay(...)
# client + settler closed automatically

Exceptions

AgentGuardError
├── PaymentBlocked           ← catch this for graceful handling
│   ├── PolicyViolation
│   └── WalletPaused
├── IdentityError
└── APIUnavailable

All exceptions are only raised when you opt into strict=True or when the API is unreachable. Default behavior is to return a receipt with approved=False and let your code branch on it.

Policy format

Policies are YAML. The SDK loads them locally for a fast fail-fast check, and sends the raw dict to the API for authoritative evaluation.

agent_id: research-agent-v1
owner_wallet: 0xAcmeCorp

spending_limits:
  per_transaction: 5          # USDC — hard cap per tx
  per_day: 50                 # rolling 24h
  per_recipient_per_day: 10   # per-recipient rolling 24h

recipient_policy:
  type: allowlist
  approved_recipients:
    - data-vendor-agent-v1
    - inference-agent-v1
    - sms-agent-v1
  fallback_action: block      # or: escalate | allow

category_policy:
  allowed: [data_purchase, compute, messaging]
  blocked: [gambling, high_risk_jurisdictions]

intent_verification:
  enabled: true
  sensitivity: medium         # low | medium | high

anomaly_detection:
  enabled: true
  alert_threshold_std_dev: 3.0

approval_rules:
  - if: transaction_amount > 2
    then: require_human_approval
  - if: recipient_is_new AND amount > 1
    then: escalate

kill_switch:
  enabled: true
  authorized_pausers:
    - 0xOperator
    - 0xSecurityLead

audit:
  log_all_checks: true        # every decision writes an on-chain receipt

Unknown keys are preserved in policy.raw and forwarded to the API, so you can iterate on the policy engine without bumping the SDK.

Framework integrations

See examples/:

• Claude Agent SDK — expose pay as an MCP tool.
• LangChain — StructuredTool the agent can call.
• AutoGen / direct use — call guard.pay() directly.
• x402 endpoints — pay any x402-protected HTTP service through the pipeline.

Development

git clone https://github.com/vikramRooT/agentguard
cd agentguard/apps/sdk/python
pip install -e ".[dev]"
pytest
ruff check agentguard tests
mypy agentguard

35 tests covering the public surface; all mocked, no network or server required.

License

MIT © 2026 AgentGuard Protocol contributors.

Links

• Repo: https://github.com/vikramRooT/agentguard
• Docs: https://github.com/vikramRooT/agentguard#readme
• Issues: https://github.com/vikramRooT/agentguard/issues
• Built for the Agentic Economy on Arc hackathon (April 2026).