Stripe MCP Server for Merchants: Setup and What to Protect

Stripe MCP is Stripe's official connection between an AI assistant and your Stripe account. It is built on MCP (Model Context Protocol), an open standard that lets AI apps use external tools and data. Once connected, an assistant can look up payments, customers and subscriptions, search Stripe's documentation, and, depending on permissions, make changes. Stripe labels the server "Public preview" in its docs, so details can change. This guide covers how to use Stripe MCP, how to add it to Claude, and what to protect.

What is the Stripe MCP server?

The Stripe MCP server gives AI agents tools to call the Stripe API and to search Stripe's knowledge base, including docs and support articles. Stripe hosts it as a remote server at mcp.stripe.com. The docs list clients such as Cursor, Claude and Codex.

In plain terms: the assistant is the MCP client, Stripe is the MCP server, and the server decides which tools the assistant can see. Which tools appear depends on what your login or key is allowed to do.

If MCP is new to you, start with the article What Is MCP? A Plain-English Guide for E-commerce Merchants.

What do you need before you start?

• A Stripe account with permission to authorize apps. Use a sandbox first (a test environment with no real money).
• An MCP-capable client: Claude (desktop app or Claude Code), Cursor, VS Code or Codex.
• Someone who can change admin settings. Administrators can switch MCP access on or off for the whole team in the Dashboard, separately for live mode and sandbox.
• If you work on a company network, check with IT. Managed networks may block external MCP servers, and the fix is to allow mcp.stripe.com.

How do you use Stripe MCP, step by step?

1. Check the admin setting. In the Stripe Dashboard, confirm MCP access is enabled for the environment you want to test. Start with sandbox only.
2. Pick the recommended setup path. Stripe's docs mark setup with a plugin as "Recommended". Open https://docs.stripe.com/mcp and follow the instructions for your client. One third-party guide says a Stripe CLI command, stripe agent setup, configures the server and keeps it updated. We could not verify that against Stripe's own docs, so check before relying on it.
3. Sign in with OAuth. Your client opens a Stripe consent page. You can grant access to one or more live accounts or sandboxes, and set different permissions per environment. Grant the sandbox first.
4. Run a read-only test. Ask something harmless: "List the last five payments in my sandbox and their statuses." Check the answer against the Dashboard.
5. Review what the assistant can see. Ask it to list its available Stripe tools. If you see write tools you did not intend, reduce permissions and reconnect.
6. Move to live mode deliberately. Authorize a live account only after the sandbox tests behave, and keep that grant read-only at first.

Stripe's official guidance is to use OAuth for interactive clients. It lets you authorize and revoke a client without sharing an API key.

How do you add Stripe MCP to Claude?

Adding Stripe MCP to Claude depends on which Claude product you use. In all cases the server address is the hosted endpoint, and the sign-in happens through Stripe's OAuth page.

Claude Code

Claude Code is Anthropic's command-line coding tool. A third-party guide shows this command:

claude mcp add --transport http stripe https://mcp.stripe.com/

Run it, then start Claude Code and complete the OAuth consent in the browser window that opens. Confirm the exact command in Stripe's docs, since tutorials age quickly.

Claude desktop app

If your plan supports adding remote servers (often called connectors), add a custom server with the URL https://mcp.stripe.com and approve access when Stripe asks. Where this setting lives and which plans include it is set by Anthropic, so check Claude's current help pages.

Cursor, VS Code and others

Stripe's docs cover Cursor and others. A third-party listing mentions a multi-client installer, npx add-mcp 'https://mcp.stripe.com', that writes the config into several tools at once. Treat any one-line installer as something to inspect before running, because it edits your tool settings.

For a similar walkthrough on a different system, see Connect Claude to Shopify: The Complete Step-by-Step Guide.

OAuth, agent API key or restricted key: which should you use?

This is where older tutorials mislead. Stripe's docs describe two methods:

Method	Best for	Notes
OAuth	Interactive use by a person in Claude, Cursor and similar	Acts on behalf of your Stripe user. Revocable without sharing a key.
Agent API key	AI apps that need independent, persistent access	A key you create with only the permissions needed. Shown with an "Agent" badge in the Dashboard.
Full-access secret key or untagged restricted key	Nothing new	Being phased out for MCP (see below).

What changed with the Stripe restricted key?

Many guides say to paste a Stripe restricted key (a key limited to chosen permissions) as a bearer token. Around late September the docs changed the wording to "agent API key". Third-party sources describe an agent key as a restricted key tagged as belonging to an autonomous agent and subject to approval rules. Read Stripe's own key documentation at https://docs.stripe.com/keys to confirm the current definition.

What is the deadline for migrating keys?

Sources conflict. The current Stripe MCP docs page says to migrate by November 30, 2026: from December, MCP stops accepting full-access secret keys and restricted keys without the Agent tag. Unsupported keys get a 401 error with an OAuth discovery challenge. Earlier versions of the page, and several third-party guides, say October 31, 2026. It looks like the date was extended, but we could not find a dated changelog confirming it. Treat November 30 as the official date per the current page, aim to migrate sooner, and re-check https://docs.stripe.com/mcp before you act.

If you use a connection that relies on an old key, replace it with OAuth or an agent API key now.

What can Stripe MCP do, and what can it not do?

Tool names vary by source and date, so check the tool table in Stripe's docs. Third-party listings describe the current server as a small set of general tools: search the API, get details on an endpoint, read through GET requests, write, create refunds, get account info, and search documentation. Earlier versions had resource-specific tools such as creating customers, payment links, invoices, and cancelling subscriptions.

Typical merchant tasks:

• Look up a customer's payments and refunds while answering a support ticket.
• Draft a payment link or invoice for review.
• Ask how a Stripe feature works, using the documentation search.
• Preview subscription metrics. Stripe said at Sessions 2026 that real-time subscription metrics were "previewed" through the Stripe MCP or Console, so expect limits.

Limits to plan around:

• It is in public preview, not generally available.
• Third-party sources say there are no event triggers, so a failed payment or dispute cannot start a workflow through MCP. That is a vendor blog's view, but it matches how MCP works: the assistant asks, the server answers.
• The assistant can be wrong. It may misread a report or pick the wrong customer. Verify numbers in the Dashboard before acting.
• Your accounting, tax and reconciliation still need the usual process.

What should you protect when you connect Stripe to an AI assistant?

Payments are the highest-risk system you can connect. A short checklist:

• Start read-only. Grant only read permissions at first, in a sandbox, then in live mode.
• Limit by environment. Use different permissions for sandbox and live in the OAuth screen.
• Avoid full-access secret keys. Never paste a secret key (starting sk_) into a chat or config file you share.
• Use agent API keys for unattended apps only, with the smallest set of permissions, and store them in a secrets manager, not in a repository.
• Keep a human in the loop for money movement. Third-party sources say Stripe asks a person to confirm certain write actions, such as refunds and outbound payments, through a link the agent returns, and that unconfirmed requests expire after 24 hours. We could not confirm the exact list of gated actions, so test it in a sandbox rather than assuming.
• Watch for prompt injection. Text in an email, a ticket or a product review can contain instructions aimed at the assistant. Do not combine Stripe write access with tools that read untrusted text without review.
• Review and revoke. Check connected apps and keys regularly, remove those you no longer use, and restrict who can enable MCP in the Dashboard.

For a wider checklist, see MCP Security: Is MCP Safe for Your Store? A Merchant Checklist.

What can go wrong?

• 401 error after December. An old secret key or untagged restricted key stops working. Switch to OAuth or an agent API key.
• Connection fails on a work network. Ask IT to allow mcp.stripe.com.
• Old tutorial commands fail. The local @stripe/mcp npm package is still published (latest listed version 0.3.3, roughly six months old when we checked). A third-party guide says it now acts as a thin proxy to the hosted server and no longer accepts the --tools flag. Even Stripe's GitHub README snippet may still show --tools=all, so docs are inconsistent. Prefer the hosted server and the official docs page.
• Missing tools. Tools shown depend on permissions. If a tool is absent, check the key or OAuth grant.
• Wrong account. By default, tools act on the account tied to your OAuth session or key. If you run multiple accounts or a platform with connected accounts, confirm which one is active. Third parties describe a separate platform key plus a Stripe-Account header for connected accounts, but this is not verified in Stripe's docs, so check there.

Is Stripe MCP worth setting up?

For support lookups, documentation questions and sandbox experiments, yes, with read-only access. For refunds or invoices, treat it as a draft tool with human approval. For comparison with other servers, see Best MCP Servers for E-commerce and Marketing Teams (2026).