Agentic commerce

PAID is the payment intelligence and orchestration platform for the agentic economy. This page explains what that means when the buyer is an AI agent — and why agent payments need different guarantees than human checkout.

Why agents break normal payment APIs

An agent moving money fails differently than a person: it retries enthusiastically, it acts on stale prices, and a leaked key lets it act at machine speed. A payments platform built for agents has to answer all three structurally — not with rate limits bolted on afterward.

How PAID answers, mechanism by mechanism

• Quote-first execution. get_quote binds amount, FX, and fee into a single quote ID; execute_payment takes that ID, so the price the agent reasoned over is the price that executes. No act-on-stale-price failure mode.
• Deny-by-default keys. Every agent key starts with no permissions; each capability is granted explicitly. A leaked key without grants can do nothing — the blast radius is the policy, not the API surface.
• Idempotent mutations. Agent retries are expected, so every mutating call is retry-safe at the contract level — no double charges from an eager loop. See architecture.
• A 24-hour sandbox, no human signup. An agent can mint a sandbox key and make its first quote call in one session, without a human filling out a form first.
• Multi-rail in one call. Cards, ACH debit, and USDC sit behind the same tools, routed deterministically — intelligent routing applies to agent payments the same as human checkout.

The wiring

The agent surface speaks Model Context Protocol over Streamable HTTP (SSE supported), with plain JSON-RPC underneath for agents that prefer raw fetch. Copy-paste configs for Claude Desktop, VS Code, and the Claude CLI — plus a live playground that mints a sandbox key in the browser — are on the for-agents page. The x402 payment-required flow is part of the same surface.

The audit trail agents inherit

Every agent-initiated payment lands in the same Ed25519-signed, hash-chained ledger as everything else — so "what did the agent spend, when, and on whose authority" is a query, not an investigation. That matters the day an operator has to answer for an agent's spending. See the hash-chained ledger.

Related

• For agents — configs, sandbox playground, tool list
• Payment orchestration — the pillar overview
• API reference · Developer docs