Let an AI agent wire up
your crypto payments
Put finality's real API inside Claude, Cursor, or Windsurf. The agent stops guessing — it reads the actual production implementation: endpoints, fields, webhook events, and the traps that are wrong the moment you guess them.
Why this exists
“Add crypto payments to my store.” Without a source of truth, an AI writes code that demos fine and fails silently in production.
- Field names are invented — plausible, nonexistent, and you find out at launch.
- Webhook event types are made up — you listen for an event that never fires, and orders never ship.
- Signature verification is written against parsed JSON instead of the raw bytes — intermittent failures, nearly impossible to reproduce.
- Endpoints and fields come from the live implementation, not an outdated version in training data.
- Integration guides name the traps outright: verify raw bytes, fulfill only on payment.settled, echo both cursor fields together.
- The agent can run a real sandbox payment end to end — instead of handing you code and hoping.
All you have to say
> Add USDC checkout to this Next.js store, test it in sandbox first
The agent reads the docs, checks whether your payout addresses are configured per chain, creates a sandbox order, simulates the payment, confirms the webhook landed — and only then hands you the integration code.
Install
Three steps, two minutes
MCP is an open protocol. Any client that speaks it works — the config shape is identical, only the file location differs.
- 1
Grab a sandbox key
Dashboard → API keys. Create one starting with sk_test_. Zero cost, never touches real money.
- 2
Add it to your MCP config
claude_desktop_config.json for Claude Desktop, .mcp.json in your project for Claude Code, or the MCP settings in Cursor and Windsurf. Same shape everywhere.
- 3
Restart the client
Then just tell the AI what you want. No npm install needed — npx fetches it.
{
"mcpServers": {
"finality": {
"command": "npx",
"args": ["-y", "@finality/mcp"],
"env": {
"FINALITY_API_KEY": "sk_test_your_key_here"
}
}
}
}It works without FINALITY_API_KEY too — only the documentation tools are available, which is enough to let an AI understand our API before you commit.
★ Always start with an sk_test_ sandbox key. Every tool reports which mode it is in, so you and the agent always know whether an order is real money.
Capabilities
28 tools
Not a mechanical wrapper over REST. Each tool carries into the agent's context when to reach for it — and what goes wrong if it doesn't.
Zero to first payment: read the docs, confirm payout addresses are set, create an order, simulate a payment end to end.
finality_docs
Search the official docs and OpenAPI spec — the agent's source of truth.
finality_integration_guide
Scenario-specific integration plans that call out that scenario's traps.
finality_create_payment
Create a payment and get the address plus checkout link.
finality_simulate_payment
Simulate a sandbox payment to exercise the full flow at no cost.
finality_payout_status
Whether each chain's payout address is configured — the first thing to check. Never echoes the address itself.
finality_x402_supported
Live list of chains and tokens supporting x402 autonomous payment, read on-chain.
Security model
This MCP cannot move your money — that isn't a policy, it's a property of the system.
finality never custodies merchant funds and never holds a private key. So “could the AI drain my wallet” doesn't arise at the architecture level — it's not that we promise not to, it's that it can't.
Refunds and withdrawals: deliberately not tools
The backend only returns an unsigned transaction — your own wallet must sign and broadcast it. Letting an agent prepare a refund is meaningless and would wrongly suggest it can move funds.
Payout addresses: API keys are hard-rejected
Changing payout addresses or managing API keys requires a logged-in session plus a second factor — an email code or a wallet signature. Calling them with an MCP key is simply refused.
Keys never enter the conversation
The key is read from an environment variable only, never accepted as a tool parameter — otherwise it would enter the conversation context and could be logged or leaked.
★ What it can do falls into three buckets: create and read payments (no funds move), read-only reporting, and documentation lookup. Creating a payment is a write, but carries zero funds risk — it only generates an address, and the customer pays into your own wallet.
Give your AI a real understanding of your payment layer
Grab a sandbox key and connect in two minutes. Nothing costs money if it fails, and nothing touches real money if it works — until you swap in a live key yourself.