Frisk

frisk-mcp

An MCP server for Frisk: screen the counterparty of an x402 payment before an agent pays it.

It exposes one tool, screen_payment, which runs Frisk's deterministic checks — address sanity, dynamic-payTo comparison, transport safety, and your own spending policy — and returns allow / review / block with reasons.

Where this belongs, and where it does not

An MCP tool runs when a model decides to call it. A payment check that the model can skip is a weaker guarantee than the same check on the code path that signs the payment.

So:

• Use the library — frisk-screen — where the money actually moves. It is MIT, dependency-free, and runs locally.
• Use this server for inspection ("is this address safe to pay?"), and for agents that cannot import a library.

Frisk is advisory in both cases: it returns a verdict and your code decides. It never holds funds and cannot stop a payment by itself.

Install

npx frisk-mcp

In any MCP client's server configuration:

{
  "mcpServers": {
    "frisk": {
      "command": "npx",
      "args": ["-y", "frisk-mcp"]
    }
  }
}

No API key, no account, no network call: with no configuration the server screens entirely on your machine.

Hosted mode (optional)

Set FRISK_API_KEY to have the same tool answered by the hosted service, which adds reputation-graph signals and returns higher-confidence verdicts:

{
  "mcpServers": {
    "frisk": {
      "command": "npx",
      "args": ["-y", "frisk-mcp"],
      "env": { "FRISK_API_KEY": "..." }
    }
  }
}

The hosted tier is in early access — email [email protected]. Everything above works without it.

variable	default	meaning
FRISK_API_KEY	unset	Enables hosted screening. Unset means fully offline.
FRISK_BASE_URL	https://api.tryfrisk.dev	Override the hosted endpoint.

The tool

screen_payment

argument	type	notes
counterparty	string, required	The address the agent intends to pay.
endpoint	string	URL that quoted the payment; checked for plaintext transport.
amount	number	Amount about to be paid.
asset	string	Asset symbol, e.g. USDC.
observedPayTo	string	The payTo the endpoint actually returned. If it differs from counterparty, that is the dynamic-payTo swap.
strictness	number 0–1	0 permissive, 1 paranoid. Default 0.3.
maxPerCall	number	Spending ceiling for one call.
allowedAssets	string[]	Assets the agent may pay in.

Returns verdict, trustScore, confidence, reasons, policyHits, source (lite or hosted), and allowed.

What it does not do

It does not detect Sybil clusters, and offline it has no view of an address's history — lite screening always reports "low" confidence for that reason. It also cannot repair the protocol underneath a payment: replay, settlement races, and facilitator trust live in x402 itself, not in the request being signed. Those limits are discussed in Five ways an x402 payment can go wrong.

License

MIT