You are connecting an AI agent (Claude Desktop, Claude Code, Cursor, or another MCP-compatible host) to BlindPay using the official MCP server, with guardrails appropriate for tools that move money.
Before configuring anything, read these sources and follow them over any prior knowledge:
- https://blindpay.com/docs/llms.txt (read the build with AI page)
- The MCP server card: https://blindpay.com/.well-known/mcp/server-card.json
Set up:
1. Install the MCP server in my host's config: command npx, args ["-y", "@blindpay/mcp"], env BLINDPAY_API_KEY and BLINDPAY_INSTANCE_ID. Use a DEVELOPMENT instance ID and key first; document exactly which config file this lands in for my host.
2. Add the agent skills knowledge layer: npx skills add blindpaylabs/skills, so the agent knows rails, corridors, fees, and KYC patterns instead of guessing.
3. Verify the connection: have the agent list its BlindPay tools, then run a read-only call (for example listing customers) before anything that writes.
4. Guardrails, written into my agent's system prompt or host permissions:
- Read operations are fine unprompted; any operation that creates a payment (quotes are fine, payouts/payins/transfers are not) requires explicit human confirmation with the amount and destination restated.
- The agent must never echo the API key, and must refuse to switch to a production instance ID without me changing the config myself.
- On a development instance, funds are test-only (USDB on testnets); state this in responses so nobody mistakes a test payout for a real one.
5. Produce a short runbook: how to rotate the key (delete and recreate in the dashboard; keys are shown once), how to point at production later, and which operations the agent is allowed to run at each stage.
Deliverables: the host config snippet, the skills install, a verified tool-listing transcript, the guardrail prompt text, and the runbook.