Stripe MCP server: scoped access for every agent

Look twice at stripe_api_write: one mode covers creating a customer and refunding a charge. Source: Stripe’s tool table, September 2026.

Why give an AI agent Stripe MCP access

Without MCP you copy a subscription ID, an invoice status and a webhook payload from the Dashboard into the chat. With stripe_api_read the agent in Claude Code looks up the customer and their last invoice itself while it fixes billing code. And stripe_api_details gives it real method parameters instead of guesses.

For a solo SaaS, stripe_analytics answers MRR and churn with Stripe’s own definitions. Stripe warns that a metric the agent calculates from raw data can differ, so ask for the tool, not the sum. That’s the Stripe AI agent integration people actually use, often ending in Google Sheets MCP. On Paddle MCP instead? See all AIPx integrations.

How to connect Stripe to Claude

mcp.stripe.com is a remote MCP server; direct setup per client is in the official Stripe MCP documentation. With AIPx as your MCP gateway, you sign in to Stripe once per environment, and live and sandboxes share one URL. Your secret key isn’t copied anywhere.

Step 1: connect Stripe live mode and a sandbox in AIPx

1. In dash.aipx.run, open Integrations → Stripe and finish the MCP OAuth flow.
2. On Stripe’s consent page, pick the sandbox first.
3. Connect Stripe again and pick live mode. It shows up as a second account of the same integration.
4. Create an MCP agent, turn on Stripe, set modes per account, save.
5. Copy the endpoint: https://dash.aipx.run/mcp/<agent>.

There is no separate Stripe Sandbox integration and no restricted-key login: each environment is one OAuth account. Sandbox goes first on purpose. If you misclick an environment, better the one where payments aren’t processed.

Stripe MCP in Claude Code

claude mcp add --transport http --scope user aip https://dash.aipx.run/mcp/<agent>

Then /mcp, pick aip, authorize, and confirm with claude mcp list. More in Claude Code MCP setup.

Connect Stripe MCP to Cursor

In ~/.cursor/mcp.json:

{
  "mcpServers": {
    "aip": { "url": "https://dash.aipx.run/mcp/<agent>" }
  }
}

See Cursor MCP.

Connect Stripe MCP to Codex

codex mcp add aip --url https://dash.aipx.run/mcp/<agent>
codex mcp login aip

See Codex MCP.

Connect Stripe to Claude on the web or desktop

Customize → Connectors → Add custom connector, paste the agent URL. Stripe has its own connector in Claude’s directory, but Claude Free allows one custom connector, and one AIPx URL carries Stripe with your other accounts. See Claude MCP connectors and Claude Desktop MCP. Stripe also ships a ChatGPT plugin; the gateway route is in ChatGPT MCP.

Direct Stripe MCP vs through AIPx

Honestly, if you need Stripe only in Claude Code and only in a sandbox, claude mcp add --transport http stripe https://mcp.stripe.com/ is simpler. The MCP server catalog shows what else fits behind one URL.

Scope Stripe MCP per agent and per environment

Stripe’s consent screen decides which environment and data an account reaches. AIPx modes decide what each agent sees there, never more than Stripe granted. See per-tool MCP permissions.

Read by column: free writes where payments aren’t processed, Confirm on live, and Cursor only looks.

The weak spot is granularity. Stripe routes every write through stripe_api_write, so AIPx can’t allow POST /v1/customers and block POST /v1/refunds. Confirm is per tool. In the AIPx chat, the Confirm card shows the Input (API method and arguments) before you press Approve. Read it. Stripe itself warns about prompt injection when Stripe MCP runs next to other servers (MCP security).

Refunds and outbound payments get a second gate: Stripe gives the agent an approval URL, you review and click Approve. Then tell the agent to retry. It won’t on its own. Unconfirmed actions expire after 24 hours.

But in Claude Code, Cursor and Codex, Confirm is a hint the client enforces, not a pause on the AIPx server. Keep auto-run off on live agents and Hide anything risky (AI agent security).

Live mode and sandboxes as separate Stripe accounts

A Stripe account can have up to five sandboxes. Live mode plus two sandboxes use exactly the 3 accounts of AIPx Free. Different rules for Claude Code and Cursor need a second agent, which means Pro at $3/month (AIPx pricing). Two unrelated Stripe accounts, say your SaaS and a client’s, connect the same way, as accounts of one Stripe integration.

Test billing changes in a sandbox before live mode

Say you’re fixing plan upgrades. The agent in Claude Code has stripe_api_write on Allow for the sandbox, so it creates a test clock, a product, a price and a subscription itself, then checks the invoice preview. Nobody clicks through the Dashboard. A webhook error goes to Sentry MCP, the follow-up to Linear MCP, upgrade conversion to PostHog MCP. On live, the same agent only reads.

One thing people forget: when you retire an environment, revoke the AIPx session under OAuth sessions in the Stripe Dashboard. Session management applies only to the environment you’re viewing, so live and each sandbox are revoked separately.