x402-trinity-mcp
Enables AI agents to check paywalled resource prices, pay and fetch content within enforced budgets, and inspect wallet status.
README
x402-trinity
A drop-in fetch replacement that lets an AI agent pay for things by itself — with hard
spending limits, no hosted wallet service, and no changes to the agent's own code.
Zero runtime dependencies. 7 KB gzipped. Buyer and seller in TypeScript; buyer in Python.
npm install x402-trinity
Why
When an agent hits a paid resource it gets 402 Payment Required. Without something to
handle that, the request simply fails.
x402-trinity handles it: reads the challenge, checks it against limits you set, signs, retries. All of it locally — no hosted wallet service, no third-party API, and the key never leaves your process.
Gas exists but the payer doesn't pay it — under EIP-3009 the facilitator submits the transfer, so an agent's wallet only ever needs USDC.
Buy (agent pays for things)
import { createX402Fetch } from 'x402-trinity';
const x402Fetch = createX402Fetch({
privateKey: process.env.X402_PRIVATE_KEY, // never hardcode
policy: {
maxAmountPerRequest: '5000', // 0.005 USDC max per call (REQUIRED)
totalBudget: '1000000', // 1.00 USDC lifetime (REQUIRED)
allowHosts: ['api.example.com'],
allowPayTo: ['0x...'],
},
});
const r = await x402Fetch('https://api.example.com/data'); // 402 handled, returns 200
Or patch the global so unmodified code pays automatically:
import { installX402 } from 'x402-trinity';
const uninstall = installX402(cfg); // globalThis.fetch now pays 402s
Confirm which wallet will pay before funding anything:
X402_PRIVATE_KEY=0x... npx x402-trinity-whoami
It prints the address and its balance on every chain, and never prints the key. If the address is not the wallet you meant, stop before sending anything.
Sell (charge for a resource, get paid)
import { createX402Seller } from 'x402-trinity/seller';
import { createFileNonceStore } from 'x402-trinity/budget-file';
const seller = createX402Seller({
payTo: '0xYourWallet', // 100% of every payment lands here
price: '10000', // 0.01 USDC, atomic units
network: 'base',
facilitator: 'https://your-facilitator.example', // REQUIRED — no default exists
nonceStore: createFileNonceStore('./.x402-nonces.json'), // REQUIRED — see below
});
// in any fetch-style handler:
const gate = await seller.guard(request);
if (gate.response) return gate.response; // unpaid or refused — hand back the 402
return new Response(yourData, { headers: seller.receiptHeader(gate.settlement) });
Both required fields are deliberate. There is no default facilitator because settling is someone's real money and guessing an endpoint is not a default. And the replay guard has to outlive the process: an in-memory one forgets every settled payment on restart, so a buyer could re-present a spent authorization and get the resource again for free. The constructor throws rather than let either be implicit.
The seller never holds a private key and never touches funds. It quotes a price and asks a facilitator to verify and settle; money moves buyer → you directly on-chain.
This release ships Base + USDC only. Any other EVM chain works through
customChains; your address is the same on all of them.
What it does
| Protocols | x402 v1 and v2, detected per response. Unknown versions and MPP challenges declined with a clear reason, never guessed |
| Chains | Base mainnet + USDC, shipped as the default. Any other EVM chain via customChains |
| Networks | short names and CAIP-2 (eip155:8453) |
| Custody | privateKey, additive shards, or remoteSign (HSM/MPC) — your choice, not the architecture's |
| Speed | 3.2 µs warm / 1.0 ms cold (TypeScript); 1.2 µs / 1.7 ms (Python) |
| Runtime | auto-detects Cloudflare Workers and changes strategy; also Node, Bun, Deno |
| Safety | mandatory caps, allowlists, never-pay-twice reconciliation, timing-hardened signing |
Another chain
Base is the default. The signing is chain-agnostic, so add whatever you need — including a testnet to rehearse on:
const fetch2 = createX402Fetch({
privateKey: process.env.X402_PRIVATE_KEY,
maxAmountPerRequest: '10000',
totalBudget: '100000',
customChains: {
'base-sepolia': { id: 84532, asset: '0x036cbd53842c5426634e7929541ec2318f3dcf7e', name: 'USDC', version: '2' },
},
});
Verify the entry against the deployed contract first: call DOMAIN_SEPARATOR() and check it
equals what this library computes. A wrong name or version produces a signature that looks
valid and the contract rejects.
It deliberately will not
Broadcast to a chain · hold funds · need gas or an RPC · take a cut of a payment (structurally impossible — EIP-3009 has one recipient) · require a hosted signer · pay without limits · guess at a protocol it doesn't speak.
Safety — read this before real money
Caps are mandatory. maxAmountPerRequest and totalBudget have no defaults; the wrapper
refuses to construct without them. An auto-payer without limits is a money leak controlled by
whoever runs the server.
totalBudget alone is per-instance. It is an in-memory counter that resets on restart, on
a new client, and on every Cloudflare Worker isolate — which is per request. On mainnet that
turns a lifetime cap into a per-request cap. Mainnet therefore requires a durable
budgetStore, or an explicit acknowledgeEphemeralBudget: true:
import { createFileBudgetStore } from 'x402-trinity/budget-file';
createX402Fetch({
...,
budgetStore: createFileBudgetStore('./.x402-budget.json'), // survives restarts
});
The key is in the process. That is the trade for having no hosted signer. Bound it:
use a dedicated wallet holding only what you would accept losing, set allowPayTo so an
exfiltrated key still cannot pay a stranger through this wrapper, and use remoteSign if you
need enclave-grade custody.
Timing hardening is hardening, not a proof. Secret scalars use blinding plus always-add-and-double, which cut the timing spread from 99.6% to 12% (Python: 99.9% → 1.1%). BigInt arithmetic is itself variable-time and no pure-JS implementation removes that.
MCP server — give an agent a wallet
MCP is how an assistant gets tools. This exposes three:
| tool | |
|---|---|
check_price |
what a resource costs, without paying |
pay_and_fetch |
fetch it, paying if it is within your limits |
wallet_status |
address, balance, spent, remaining |
{
"mcpServers": {
"x402-trinity": {
"command": "npx",
"args": ["-y", "x402-trinity-mcp"],
"env": {
"X402_PRIVATE_KEY": "0x...",
"X402_MAX_PER_REQUEST": "50000",
"X402_TOTAL_BUDGET": "1000000",
"X402_BUDGET_FILE": "./.x402-budget.json",
"X402_ALLOW_HOSTS": "api.example.com",
"X402_NETWORKS": "base"
}
}
}
}
A model decides when these run. It cannot be reasoned with about budgets, and a
paywalled page can claim any price it likes. So the limits are not parameters the model can
set — they come from the environment, and the server refuses to start without
X402_PRIVATE_KEY, X402_MAX_PER_REQUEST and X402_TOTAL_BUDGET.
Set X402_BUDGET_FILE too, or the lifetime ceiling resets every restart. Set
X402_ALLOW_HOSTS and nothing else can be paid, however convincing the challenge looks.
Use a wallet holding only what you would accept losing.
Python
Same protocol, standard library only. Buyer only — to charge for a resource, use the TypeScript seller. Aimed at long-lived processes: on-body agents, robotics controllers, harvesting scripts.
from x402_trinity import X402Client, Policy
client = X402Client(
private_key=os.environ["X402_PRIVATE_KEY"],
policy=Policy(max_amount_per_request=5000, total_budget=1_000_000,
allow_hosts=["api.example.com"]),
)
body = client.urlopen("https://api.example.com/data").read()
Or as a decorator, so payment-unaware code just works:
from x402_trinity import x402_telemetry
@x402_telemetry(private_key=KEY, policy=Policy(...))
def harvest():
return urllib.request.urlopen("https://sensor.local/v1/lidar").read()
The warm path is 1.1 µs — a long-lived controller is warm after its first payment.
Development
npm install
npm run build # dist/*.js + *.min.js + *.d.ts
esbuild, typescript and wrangler are dev dependencies only, used for building.
The shipped package has zero runtime dependencies.
License
MIT. This software moves money — see LICENSE for the additional notice, and set your spending caps.
Recommended Servers
playwright-mcp
A Model Context Protocol server that enables LLMs to interact with web pages through structured accessibility snapshots without requiring vision models or screenshots.
Audiense Insights MCP Server
Enables interaction with Audiense Insights accounts via the Model Context Protocol, facilitating the extraction and analysis of marketing insights and audience data including demographics, behavior, and influencer engagement.
Magic Component Platform (MCP)
An AI-powered tool that generates modern UI components from natural language descriptions, integrating with popular IDEs to streamline UI development workflow.
VeyraX MCP
Single MCP tool to connect all your favorite tools: Gmail, Calendar and 40 more.
graphlit-mcp-server
The Model Context Protocol (MCP) Server enables integration between MCP clients and the Graphlit service. Ingest anything from Slack to Gmail to podcast feeds, in addition to web crawling, into a Graphlit project - and then retrieve relevant contents from the MCP client.
Kagi MCP Server
An MCP server that integrates Kagi search capabilities with Claude AI, enabling Claude to perform real-time web searches when answering questions that require up-to-date information.
Neon Database
MCP server for interacting with Neon Management API and databases
Exa Search
A Model Context Protocol (MCP) server lets AI assistants like Claude use the Exa AI Search API for web searches. This setup allows AI models to get real-time web information in a safe and controlled way.
Qdrant Server
This repository is an example of how to create a MCP server for Qdrant, a vector search engine.
E2B
Using MCP to run code via e2b.