--- url: /docs/mcp.md description: >- The BlindPay MCP server lets AI agents quote, send, and track stablecoin payments. Connect Claude, ChatGPT, Cursor, Codex, VS Code, or any MCP client over OAuth. --- The BlindPay [Model Context Protocol (MCP)](https://modelcontextprotocol.io) server is a hosted endpoint that gives your AI agent the BlindPay API as tools. Ask in plain English and the agent can quote a payout, create a customer, issue a virtual account, or check where a transfer is, without leaving the chat or the editor. It works with any client that supports remote MCP servers, including Claude, ChatGPT, Claude Code, Cursor, Codex, VS Code, and Grok. You sign in with your BlindPay account over OAuth, so no API key goes into the client. ## Server URLs | URL | What it serves | | --- | --- | | `https://mcp.blindpay.com/mcp` | Every tool, with OAuth sign-in. Use this one by default. | | `https://mcp.blindpay.com/mcp/readonly` | Only the read tools, for agents that should look but never change anything. | | `npx -y @blindpay/mcp` | The same tools as a local server, authenticated with an API key and instance ID from your environment. | ## Set up your client Pick your client for step-by-step setup: * [BlindPay MCP for Claude](/docs/mcp/claude-connector): Claude on the web and the Claude desktop app, through a custom connector. * [BlindPay MCP for ChatGPT](/docs/mcp/chatgpt): a connector in ChatGPT developer mode. * [BlindPay MCP for VS Code](/docs/mcp/vscode): one-click install, or `.vscode/mcp.json`. * [BlindPay MCP for Claude Code](/docs/integrations/claude-code): `claude mcp add` from the terminal. * [BlindPay MCP for Cursor](/docs/integrations/cursor): `.cursor/mcp.json` with BlindPay's public OAuth client. * [BlindPay MCP for Codex](/docs/integrations/codex): `codex mcp add` for the CLI, IDE extension, and desktop app. * [BlindPay MCP for Grok](/docs/integrations/grok): a custom connector on grok.com, or Grok Build in the terminal. * [BlindPay MCP for OpenClaw](/docs/integrations/openclaw): the autonomous agent runs payouts over MCP. * [BlindPay MCP for Hermes Agent](/docs/integrations/hermes): the Nous Research agent, connected over MCP. Using another client? Add `https://mcp.blindpay.com/mcp` as a remote (Streamable HTTP) server and sign in when it asks. ## What can your agent do with it? The tools are generated from the public API, so most things you can do with the API, an agent can ask for. Every tool is listed in the [BlindPay MCP tools reference](/docs/mcp/tools). * Create customers, run KYC and KYB, and answer requests for information * Quote payouts, payins, transfers, and FX rates before anything moves * Send payouts from EVM chains, Solana, and Stellar to bank accounts, then track them * Issue virtual accounts that receive USD bank transfers and settle to stablecoins * Add bank accounts and blockchain wallets, and read wallet balances * Register payables, set partner fees, and manage webhook endpoints * Look up available rails, required bank details, SWIFT codes, and NAICS codes ```text [Prompt] Use the BlindPay tools to create a payout quote for 1000 USDC to a USD ACH bank account. Show me the fees and rate, then execute after I confirm. ``` ## How does sign-in work? When the client connects, it opens a BlindPay sign-in page. You sign in with your BlindPay account and pick the one instance the agent may use. The server fills that instance into every call, so the agent cannot reach another one, and you can disconnect the agent at any time from **Settings**, then **Connected apps**, in the BlindPay dashboard. When it connects, the server also sends the agent instructions: the call order for payouts and payins, amount units (cents for quotes and payouts, whole tokens for wallet balances), and the steps a person must do in the dashboard. Webhook events and their payload schemas are exposed as resources under `blindpay://webhooks/events`. ## Is it safe to let an agent move money? Every tool call goes through the same API as a direct request, with the same KYC and KYB requirements, limits, and audit trail. Every tool also carries MCP annotations (read-only, destructive), so clients that support them can ask before running anything that is not a read. The server does not block a call on its own. Keep a confirmation step in your agent prompt or app flow, start on a development instance, or connect to `/mcp/readonly` until you are ready. ## How do I test without real money? Connect the agent to a development instance. KYC is approved automatically, payins and payouts complete on their own, and on-chain flows use USDB, a test token with no real funds behind it. A `request_amount` of `66600` makes a payout or payin fail and `77700` makes it refunded, so the agent can practice the unhappy paths too. See [Sandbox vs. production](/docs/learn/sandbox-vs-production). ## Next steps * [BlindPay MCP tools reference](/docs/mcp/tools) * [Build with AI](/docs/build-with-ai): Agent Skills, the CLI, and prompts for AI app builders * [API reference](https://blindpay.com/docs/api/reference)