Connect Stripe, PayPal, Square, or a bank like Qonto to Claude or Cursor through MCP and you’ve opened a new path into whatever falls inside your cardholder data environment. PCI DSS doesn’t mention AI clients or tool calls — it was written for POS terminals and payment gateways — but the requirements don’t care what’s on the other end of the connection. This post maps the parts of PCI DSS that actually bite once a model can call a payment tool. It is not legal or compliance advice; it is the architecture questions to ask before you connect one.

The first question: does the tool touch cardholder data at all

Most payment MCP servers are built to avoid putting raw card data anywhere near an AI client. A Stripe tool that creates a charge from a customer ID, or a Square tool that looks up an order total, deals in tokens and references — the primary account number and the CVV already live behind the processor’s own vault and never pass through the tool call at all. That’s the easy case, and it’s the one worth checking for first: read the actual tool schema, not the vendor’s marketing copy, to confirm a result never includes a full PAN, expiry, or CVV field.

The harder case is a tool whose output legitimately needs more than a token — a dispute-resolution tool returning the last four digits and a cardholder name together, or a legacy integration that wasn’t built with an AI client in mind. The moment real cardholder data appears in a tool result, that result is inside your cardholder data environment the instant it lands in the model’s context, and every requirement below applies to it.

The requirements that actually touch an MCP setup

Requirement 7 — restrict access by business need to know. A payment MCP server’s OAuth grant is usually all-or-nothing: the scope that lets a support tool look up a transaction is the same scope that lets it issue a refund. PCI DSS expects access scoped to what a role actually needs, not to what the vendor’s API happens to bundle together. That’s a per-tool allow, ask, or block rule in front of the server, not a single approve-or-deny click on connection — see access control for MCP teams.

Requirement 8 — identify and authenticate access. A shared API key for a payment server, pasted into every teammate’s claude_desktop_config.json or .cursor/mcp.json, means every refund or lookup is attributed to the same identity. Per-user OAuth with individual grants is what this requirement is actually asking for; a static secret shared across a team is the thing it exists to rule out.

Requirement 10 — log and monitor all access. This is the one that trips up AI tool calls specifically. A chat transcript sitting on one person’s laptop is not a log a PCI assessor can review, and it disappears the moment that machine is wiped. A central record of every call — who, which client, which server, which tool, the arguments, the outcome — is what the requirement expects. What that log needs to contain is covered in the MCP observability post.

Requirement 12 — manage service providers. Every payment MCP server you connect is a service provider the moment it can touch cardholder data or the systems around it, whether anyone ran it through vendor review or not. PCI DSS expects a list of those providers, evidence they were checked, and a way to re-check them. See how to vet an MCP server and why a private registry beats an unmanaged pile of config files.

Tokenization doesn’t make the tool call disappear

Most of the catalog’s payment servers lean on tokenization precisely so an AI client never needs to see a real card number — Stripe, PayPal, and Square all work this way by default. That keeps the data itself out of scope, but it doesn’t make the tool call invisible: a refund, a capture, or a void is still a real action against a real transaction, and an agent that can call it needs the same approval discipline as a person clicking the same button in a dashboard. A human-in-the-loop approval step on anything that moves money is a control PCI DSS doesn’t name directly but every assessor will ask about once they understand what the tool can do.

What an assessor will actually ask for

  • Read the schema of every tool on a connected payment server and confirm what’s actually in its inputs and outputs — not what the vendor’s docs claim.
  • Pull one real tool call from the log and show who made it, from which client, with which arguments, and what happened.
  • Show that a departed employee’s access to a payment server was revoked the same day, not whenever someone remembered to rotate a shared key.
  • Walk through what happens if a connected payment server is suspected of being compromised — see what to do when an MCP server might be compromised.
Where gate fits, and where it doesn’t: gate scans a payment server’s tool descriptions and schemas before anyone connects it, sets per-tool allow, ask, or block rules per team and per AI client instead of one all-or-nothing OAuth grant, and logs every call with an integrity-hashed, exportable audit trail. That covers the access-restriction and logging requirements above. gate does not itself hold cardholder data, does not inspect the payment processor’s own infrastructure, and connecting through it does not make your organization PCI DSS compliant on its own — that still depends on your cardholder data environment, your QSA’s scope, and controls gate has no visibility into. Start from the security overview or run a server through the scanner before it touches a live payment account.

A sequence that works before the assessment

  1. Inventory every payment-adjacent MCP server anyone has connected, personal laptops included.
  2. Read each one’s tool schemas and confirm whether real cardholder data ever appears in a result.
  3. Replace shared keys with per-user OAuth, scoped per role, wherever the server supports it.
  4. Put money-moving tools behind a human approval step, not implicit trust in the model.
  5. Route every call through one point that logs centrally, and export that log before the assessor asks for it.

None of this is unique to AI. PCI DSS has always asked the same three questions — who can touch cardholder data, is it logged, is the vendor vetted — of every new integration a payments team ships. MCP just adds a caller that decides which tool to use mid-conversation instead of a person clicking a known button, and most assessors haven’t seen that caller yet. A gateway in front of your payment servers is what gives that caller the same visibility every other integration in your cardholder data environment already has.