AI AgentA program capable of making decisions and acting autonomously to achieve a given objective. Unlike a chatbot, an agent can perform actions in the real world: send emails, call APIs, trigger payments.API (Application Programming Interface)An interface that allows two programs to communicate. The AgentGate API allows your AI agent to send payment requests and receive authorization decisions.API KeyA secret identifier that authenticates your agent with AgentGate. Each agent has its own key. It must be kept secret and stored in environment variables, never in source code.Payment IntentA request submitted by an AI agent to make a payment. It contains the amount, beneficiary, category, and reason. AgentGate analyzes this intent before approving or rejecting it.MCP (Model Context Protocol)A standard created by Anthropic that allows AI agents to connect to external tools in a standardized way. AgentGate exposes an MCP server, allowing Claude and compatible agents to use it without custom integration.STANDARD ModeAn agent operating mode in which it follows the organization's global rules. Suited to general-purpose agents that must comply with common policies.CUSTOM ModeAn agent operating mode in which it follows rules specific to that agent. Suited to agents dedicated to a precise task with differentiated permissions.PolicyA set of rules that define what an agent can or cannot pay. Examples: maximum amount, authorized beneficiaries, categories, time window, daily cumulative limit.SandboxA test mode in which payments are simulated without being actually executed. Allows you to validate your rules and integration without financial risk.Audit TrailA chronological and tamper-proof log of all events related to a transaction: submission, decision, approval, execution. Each event is cryptographically hashed to guarantee its integrity.BeneficiaryThe entity that receives the payment: a person, a company, a service. AgentGate allows you to define a whitelist of authorized beneficiaries.WhitelistA list of pre-approved beneficiaries or categories. A payment to a beneficiary not on the whitelist can be automatically blocked or sent for human approval.Human-in-the-loopA process by which a human must approve or reject a payment before its execution. AgentGate notifies approvers by email, Slack, Telegram, or Discord.Prompt InjectionAn attack technique where malicious text in data processed by an agent instructs it to perform undesired actions, such as triggering a fraudulent payment.WebhookA URL to which AgentGate sends automatic notifications upon events (payment submitted, decision made, etc.). Used for integration with Slack and Discord.