AgentGuard® Score: payment risk for merchants and AI agents

Payment risk check

For a business that takes card payments, or an AI agent that sells on its behalf. Answer from the last 30 days. The first four questions are required.

Industry

Card payments a month

Chargebacks a month

Of those, fraud claims

Where you sell

Time taking payments

Fraud tools you use 3D Secure Address checks (AVS) CVV checks Velocity limits Device fingerprinting Stripe Radar Manual review

Leave all unchecked if you use none.

Documents you have ready Terms of service Privacy policy Refund policy Business verification PCI compliance Processing history

Leave all unchecked if none are ready.

Your answers go to our hosted scoring service to compute the result. We ask for no name or email. The privacy policy says what is kept.

Score an AI agent that moves money

Five questions about the controls around an agent that pays or gets paid: who is accountable, where its money sits, its limits and its records. The score covers what you report. It verifies nothing.

What kind of agent is this?

Where does the agent's money sit?

Is a named human accountable for this agent, and can a third party verify who that is? Yes No

Are per-transaction and daily limits enforced somewhere the agent cannot change them? Yes No

Is every payment action logged with a timestamp the agent cannot edit? Yes No

The five answers go to our hosted scoring service. No email. No report link is created.

API for developers and agents

Both checks are plain HTTPS endpoints on agentguard.run. No key. JSON in, JSON out. Requests are rate limited per address.

Payment risk check

Send industry, monthly_transactions, total_disputes and fraud_disputes. Optional: region, business_age, avg_transaction_value, refund_rate, customer_mfa, checkout_platform, and the lists fraud_tools, current_psp and compliance_readiness. Values use the option ids from the form above. Any other field is refused.

curl -s https://agentguard.run/api/payment-risk \
  -H 'Content-Type: application/json' \
  -d '{"answers":{"industry":"saas","monthly_transactions":"500-2000","total_disputes":"1-5","fraud_disputes":"0","region":"US","fraud_tools":["3ds","cvv"]}}'

The reply carries score (0 to 100), band_label, approval_likelihood (high, moderate or low), reserve_estimate, recommended_providers, flags, recommendations, the dispute ratio in vamp, and a disclaimer.

AgentGuard Score for an AI agent

curl -s https://agentguard.run/api/score \
  -H 'Content-Type: application/json' \
  -d '{"answers":{"agent_type":"Customer service agent","human_sponsor":true,"wallet_type":"None yet","transaction_limits":true,"audit_trail":true}}'

Every one of the five answers is required. The reply carries score, tier, a category breakdown and factors with recommendations.

Inside Claude Code and Codex

The AgentGuard plugin includes the agentguard-score skill and the agent_score MCP tool. Your agent asks the five questions, names the service it will send them to, and sends them only after you say yes.