com.stripe/mcp

The Stripe MCP server lets an AI agent call the Stripe API and search Stripe's documentation and support articles. Stripe hosts it at https://mcp.stripe.com and signs you in with OAuth, or you can use an agent API key. In Claude Code: claude mcp add --transport http stripe https://mcp.stripe.com/. Use a sandbox first, because the same tools can write to a live account. We checked Stripe's docs on 31 October 2026.

Quick facts

What it lets an agent do

Stripe's tool list is short on purpose. Instead of one tool per API method, it exposes a search-and-call pattern, which the docs say makes much of the API available without bloating the context window.

The supported API methods listed in the docs include customers, charges, refunds, PaymentIntents, PaymentMethods, SetupIntents, Checkout Sessions, invoices and invoice items, credit notes, subscriptions, schedules, coupons and more. Some are flagged human confirmation, for example creating a refund.

For a developer building on Stripe, the everyday uses are: look up how to do something and get the right parameters, check what happened to a payment or subscription in a sandbox, create test objects, and have the agent plan an integration. The implementation planner is meant to walk you through Stripe products for accepting payments, selling online, billing or other integrations.

Install it in each agent

The recommended route: the Stripe CLI

Stripe recommends agent plugins, which configure the server, install its skills and keep them up to date:

npm install -g @stripe/cli@latest
stripe agent setup

The command detects which agents you use and runs the applicable setup. The stripe/ai repo also lists plugin commands per agent:

Manual skills install: npx skills add https://docs.stripe.com. The repo notes manually installed skills do not auto-update; run npx skills update -y.

Manual setup

Claude Code

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

Start Claude Code and run /mcp to authenticate with OAuth. For an agent API key instead, put the server in .mcp.json and reference an environment variable, not the key itself:

{
  "mcpServers": {
    "stripe": {
      "type": "http",
      "url": "https://mcp.stripe.com",
      "headers": {
        "Authorization": "Bearer ${AGENT_API_KEY}"
      }
    }
  }
}

Cursor, in ~/.cursor/mcp.json:

{
  "mcpServers": {
    "stripe": { "url": "https://mcp.stripe.com" }
  }
}

Codex

codex mcp add stripe --url https://mcp.stripe.com

For a key, in config.toml:

[mcp_servers.stripe]
url = "https://mcp.stripe.com"
bearer_token_env_var = "AGENT_API_KEY"

Replace AGENT_API_KEY with the name of your environment variable, not the name wrapped in braces. The docs note that the ChatGPT desktop app, Codex CLI and the Codex IDE extension share the same MCP configuration on your machine.

VS Code, in .vscode/mcp.json:

{
  "servers": {
    "stripe": { "type": "http", "url": "https://mcp.stripe.com" }
  }
}

Gemini CLI. Stripe's page has no Gemini tab. The generic instruction is to use https://mcp.stripe.com as the server URL with OAuth where supported. In ~/.gemini/settings.json that would be {"mcpServers":{"stripe":{"httpUrl":"https://mcp.stripe.com"}}}, with /mcp auth stripe to sign in. That is our adaptation, and we did not test it in Gemini CLI.

Claude Desktop and ChatGPT have an official Stripe connector and plugin respectively; for Team and Enterprise accounts in Claude, an owner must add the connector to the workspace first.

Authentication, access and the November 2026 change

Stripe's server supports two authentication methods.

1. OAuth when a client accesses Stripe on behalf of your Stripe user. The consent page lets you grant access to one or more live mode accounts or sandboxes and set different permissions for each.
2. Agent API keys when an AI application needs independent, persistent access. Create one with only the permissions the client needs. They show an Agent badge in the Dashboard.

Update by 30 November 2026. Stripe's docs announce that starting in December, the server will stop accepting full-access secret keys and restricted API keys that do not have the Agent tag. Requests with an unsupported key get a 401 response with an OAuth discovery challenge. If you pass a normal key today, replace it with a new agent API key or switch to OAuth, test in a sandbox, then expire the old key.

Admin controls. An administrator can enable or disable MCP access for the whole team in Dashboard settings, separately for live mode and sandbox. Users see authorised clients under OAuth sessions in their settings and can revoke each one. Administrators with team-management permission can view or revoke other users' sessions. If your IT blocks external MCP servers, allow mcp.stripe.com.

Safety: this one touches money

Be stricter here than with a docs or search server.

• Sandbox first. Connect to a sandbox for every experiment. Stripe's OAuth page lets you choose sandboxes and live accounts separately.
• Turn on tool approval. Stripe's docs say to enable your client's tool-approval setting so you can review calls before they run, and to be careful when you use the Stripe server with other servers because of prompt injection.
• Narrow the key. An agent API key with only the permissions the task needs.
• Know which calls are write calls. stripe_api_write can reach POST, PATCH, PUT and DELETE methods. The docs mark some as human confirmation (creating a refund, and the Treasury movement tools). That confirmation is a guardrail in the product; your client's approval prompt is another layer. Keep both on.
• Treasury tools are in public preview and include send_money and transfer_money. Do not enable them for an agent that reads untrusted text.
• Customer data. Customers, charges and invoices contain personal information. It goes to your model provider.
• Revoke on departure. Remove OAuth sessions when someone leaves or a laptop is lost.

The combination to avoid is an agent that reads untrusted content (support tickets, customer notes, web pages) with live-mode write access.

Real tasks with copy-ready prompts

All of these assume a sandbox.

Learn the right call

Using the Stripe docs and API search, show me how to create a
subscription with a trial and a coupon. List the parameters and give
a curl example. Do not call anything.

Check a payment

Find the most recent failed PaymentIntent in this sandbox and explain
the failure code in plain words.

Create test data

Create a test customer, a product with a monthly price and a
subscription for that customer. Return the IDs.

Plan an integration

Use the implementation planner: I want to sell a one-off digital
download and a monthly plan on a Next.js site. Give me the Stripe
products I need and the order to build them.

Reconcile

List invoices from the last 7 days that are open or uncollectible and
summarise by customer. Read-only.

Pair with Context7 MCP for your web framework's docs and GitHub MCP for the pull request.

A test plan for payment code written with an agent

The server is most useful while you write payment code, and payment code needs more testing than most. The sandbox exists for it. A plan that uses the agent without trusting it:

1. Model the flow on paper first. Write down each state a payment can reach: created, requires action, succeeded, failed, refunded, disputed. The agent can draft this from the docs, but you approve it.
2. Build against the sandbox only. Use the agent API key and OAuth session scoped to the sandbox, with a name that makes the environment obvious.
3. Cover failure, not only success. Ask the agent to produce test data for declined cards, authentication required, an expired card and a refund. Stripe documents test values for these in its testing guide, so ask the agent to cite the page it took them from.
4. Handle webhooks explicitly. Most real payment bugs live in event handling: duplicate events, events out of order and events that arrive after your page redirected. Write a test that delivers the same event twice and one that delivers events in the wrong order.
5. Check idempotency. A retried request must not charge twice. Ask the agent to show where in your code an idempotency key is set, and confirm by reading it.
6. Reconcile. After a test run, list the sandbox objects the agent created and compare with what your database recorded. Any mismatch is a bug in your code or in the agent's.
7. Review by a person who knows payments before anything touches live mode.

The agent is good at step 1 and at writing boilerplate for steps 3 and 4. It is not a substitute for step 7, because an agent will happily write code that works in every test you thought of and fails on the one you did not.

Cleaning up after a session

Agents leave things behind. After a sandbox session, a short cleanup keeps the account understandable.

• List what was created. Ask for the customers, products, prices and subscriptions created today, with IDs.
• Delete test objects you do not need, through the Dashboard or the agent with approval on, and read each delete before you confirm.
• Revoke the session or key if the task is finished and nobody will need it. Sessions and Agent-badged keys are listed in the Dashboard.
• Record what you learned, such as the product IDs your code now uses, in the repository, not in chat.

Naming helps. Tell the agent to prefix everything it creates with a marker such as agent-test-, so a later search finds them all. A sandbox full of unnamed "Test customer 41" objects is hard to reason about.

Preparing for the key change

The authentication section says the server will stop accepting full-access secret keys and restricted keys without the Agent tag starting in December. If any script, CI job or teammate uses a plain key with the server today, it will fail then. A short checklist, to finish well before the 30 November date:

Do this in the sandbox first, then in live mode. The agent badge makes the keys easy to audit afterwards, which is a benefit beyond the deadline.

What to put in front of a human

Not every action needs the same level of care. A simple grid helps teams decide what the agent may do alone.

Write the grid into your team's page and the agent's rules file, so the agent knows the boundary and so does everyone else. See best MCP servers for where it sits in a typical setup.

Rolling it out to a team

Stripe gives administrators the controls, so use them before individuals start connecting their own clients.

1. Decide whether MCP access is on at all, and set live mode and sandbox separately in Dashboard settings. Many teams will want sandbox on and live off to begin with.
2. Name an owner for the agent API keys. Keys show an Agent badge, so you can audit them in one place.
3. Tell people which clients are approved. OAuth sessions list the client and can be revoked one by one or all at once.
4. Write down the rule for live mode: read-only prompts, a named person approves anything that writes.
5. Review the OAuth sessions list on a schedule and remove stale clients.

Limits and known issues

• Preview areas. Analytics is a preview; Treasury is a public preview. Behaviour can change.
• Key deadline. The December 2026 change breaks old keys.
• Managed networks may block mcp.stripe.com.
• Context and tool shape. The generic search-and-call design keeps the context small, and it means the agent must find the right method first. Give it a pointer to the object type you care about.
• Not a replacement for webhooks or your backend. The server is for an agent working with your account, not a production integration path.
• Gemini and some clients are not documented in Stripe's tabs, so you may need generic setup.

Alternatives

Our take

We read Stripe's documentation and repo; we did not connect an account.

The design is thoughtful. A handful of tools, a search step to find the right API method, human confirmation on the riskiest calls, admin switches for live mode and sandbox separately, and a clear path to revoke sessions. That is a good model for a vendor server that can move money.

Our rule would be sandbox for everything until a task proves it needs live mode, read-only prompts in live mode, and no agent that reads untrusted text gets write access to a live account. Use OAuth for interactive work, and agent API keys scoped to the minimum for anything that runs on its own.

Calendar the date: if you use a standard key in an MCP config today, Stripe says to migrate by 30 November 2026, because it stops being accepted from December.

Setup checklist

1. Install the Stripe CLI and run stripe agent setup, or add the server by hand.
2. Authenticate with OAuth and pick a sandbox.
3. Turn on tool approval in your agent.
4. Try a read-only prompt.
5. If you use a key, create an agent API key and store it in an environment variable.
6. Replace any old secret or restricted keys before the November 2026 deadline.
7. Review OAuth sessions in settings each quarter.

Frequently asked questions

What is the URL? https://mcp.stripe.com.

How do I install it in Claude Code? claude mcp add --transport http stripe https://mcp.stripe.com/, then /mcp.

What changes on 30 November 2026? Stripe says to migrate by that date, because old secret and restricted keys without the Agent tag stop being accepted from December.

Can the agent move money? Treasury tools exist in public preview with human confirmation, and write API calls are possible.

How do I revoke access? Under OAuth sessions in your Stripe settings.

*Product names are used descriptively. We are not affiliated with Stripe.*