Securing the Stripe MCP server

Updated

Stripe runs an official MCP server at https://mcp.stripe.com. Through it an AI agent can search and read almost anything in a Stripe account — customers, payments, invoices, subscriptions, balances — and, with one general write tool, create, update and delete through the API: refunds, cancelled subscriptions, voided invoices, coupons, webhook endpoints. Stripe’s preview tools go further and move money.

What can go wrong

One tool covers every change. Stripe exposes most writes through a single general tool (stripe_api_write) that can send any create, update or delete the API allows. Giving an agent “write” therefore gives it refunds and cancellations as much as a new coupon.

Money can move. Stripe’s Treasury tools, in public preview, send and transfer money. Stripe already asks a person to approve some actions through a link — refunds and outbound payments among them — and that approval expires.

It reads personal and payment data. Customer names, emails and addresses, payment methods, issued cards and cardholders, payouts. All of it comes back to the agent.

Records can carry instructions. Customer names, descriptions and metadata are written by your customers. Stripe itself tells users to be cautious when running other MCP servers alongside its own, because of prompt injection.

Narrow the server itself

  • An agent API key, restricted. Stripe’s MCP server accepts OAuth or an agent API key sent as a bearer token. Stripe says that from 31 October 2026 full secret keys, and restricted keys without the agent tag, are refused. Use a restricted agent key that allows only the resources the agent needs.
  • Sandbox first. Build and test against a sandbox account. Stripe lets admins turn MCP access on or off per team, separately for live mode and sandbox — keep live mode off until you are ready.
  • Keep Stripe’s own approvals on. They are a second check on the actions Stripe considers sensitive.

Govern each agent’s calls

An MCP gateway decides each call per agent, before it reaches Stripe (what is an MCP gateway?):

  • Read-only by default. Most agents only need to look things up: grant them the read and search tools, and the general write tool only to the agents whose job is to change records.
  • A person before money moves. Hold refunds, cancellations and any Treasury action for approval in your own workflow — with the approver recorded beside the call, not only in Stripe.
  • Check what goes out. Inspect what the agent sends for card numbers and credentials before it reaches any server.
  • Watch what crosses servers. Customer data an agent just read from Stripe should not flow, unexamined, into a message or another system.
  • One record of every lookup and every change, by agent.

How AtmanID approaches it

  1. Register Stripe’s server with https://mcp.stripe.com as its upstream and the protocol set to Standard MCP, and save a restricted agent key as its stored credential — encrypted, never shown again, sent only on calls AtmanID has allowed (put a hosted MCP server behind AtmanID).
  2. Review the server’s tools; a tool that later changes is held until you approve it.
  3. Grant reads broadly and the write tool narrowly. Because one tool carries every change, add an intent policy that holds that tool for a person — or denies it — for every agent that does not need to change records, and deny the money-moving preview tools outright (MCP gateway policy engine).
  4. Mark the server as handling personal data and turn on response redaction, and AtmanID removes card numbers and Social Security numbers from its responses before the agent sees them (until it is on, it counts what it would remove).
  5. Start in Shadow to see what would be refused, then Enforce. Connect agents to the server’s AtmanID endpoint (connect any MCP client).

Checklist

  • A restricted agent API key for the agents, never a full secret key
  • A sandbox account until the setup is proven; live-mode MCP access off until then
  • Read and search tools for most agents; the write tool for few
  • Refunds, cancellations and money movement held for a person (and Stripe’s own approvals left on)
  • Card numbers and other identifiers redacted from responses
  • The Stripe key is held by the gateway, not in any agent’s configuration

Related: What is an MCP gateway? · AI agent security · How to govern AI agents