A non-custodial marketplace where AI agents discover and pay for API and MCP tool calls in USDC, per call, over x402 on Base.
A provider lists a tool. An agent finds it in the registry and calls it. The tool answers
402 Payment Required with a price. The agent signs a USDC payment authorisation and
retries. The call goes through and the money moves wallet to wallet.
Fatstack never touches the money. There is no custody, no withdrawal function, no platform wallet in the payment path, and no stored private keys. Payment is a direct transfer from the agent's wallet to the provider's — or, once a provider's promotional period ends, through an immutable splitter contract that pays the provider and a 2% platform fee in one atomic transaction. If a change would let Fatstack move someone else's money, it is a bug.
Add one MCP server and your agent has the whole catalog:
https://www.fatstack.net/api/mcp
Listing what is available is free; only calls cost anything. There is no account, no API key
and no signup — the wallet is the identity. Paying needs an
x402-capable client; any MCP client can browse.
Published in the official MCP registry as net.fatstack/registry.
This repository holds the parts you run, audit, or depend on. The marketplace application itself is closed; everything an integrator touches is here, under MIT.
sequenceDiagram
participant A as Agent
participant T as Tool (provider)
participant F as Facilitator
participant C as USDC on Base
A->>T: call the tool
T-->>A: 402 Payment Required + price, asset, payTo, resource
Note over A: signs an EIP-3009 authorisation<br/>covering exactly this resource
A->>T: retry with X-PAYMENT
T->>F: verify
F-->>T: valid
T->>T: run the handler
T->>F: settle
F->>C: transferWithAuthorization
C-->>F: mined
T-->>A: 200 + result
The agent signs an authorisation, not a transaction: it never pays gas and never sends a transfer itself. The facilitator broadcasts it and pays the gas.
Payments are final. Once settled on chain the transfer cannot be reversed, and there are no refunds. That sentence appears in every 402 body, every quote object, and every payment-facing page, and it is not decoration — it is the property that makes per-call pricing work without an escrow.
Put a paywall in front of a route you already have. Two lines of setup, and the price is per call.
npm install @fatstack/x402import { Hono } from 'hono';
import { paywall } from '@fatstack/x402';
import { honoPaywall } from '@fatstack/x402/hono';
const app = new Hono();
app.use(
'/convert',
honoPaywall(
paywall({
price: '0.001', // USDC, per call
wallet: '0xYourWallet', // paid directly, wallet to wallet
toolId: 'unit-convert',
network: 'base-sepolia', // test here before base
}),
),
);
app.post('/convert', (c) => c.json({ celsius: 20, fahrenheit: 68 }));Adapters ship for Hono, Express and Next on their own entry points —
@fatstack/x402/hono, /express, /next — so importing the client into an edge runtime
does not drag a Node framework in with it.
Full walkthrough, including listing the tool: docs/provider-quickstart.md
Give a model a set of paid tools it can call, with a spending limit it cannot exceed.
npm install @fatstack/ai-sdk-tools ai viemimport { generateText } from 'ai';
import { fatstackTools } from '@fatstack/ai-sdk-tools';
import { privateKeyToAccount } from 'viem/accounts';
const tools = await fatstackTools({
wallet: privateKeyToAccount(process.env.AGENT_PRIVATE_KEY),
networks: ['base'], // the catalogue is mainnet-only; USDC here is real
guards: {
maxPerDay: 0.5, // hard ceiling per UTC day, in USD
maxPerCall: 0.01, // refuse any single call dearer than this
},
});
const { text } = await generateText({ model, tools, prompt: 'Convert 20C to F.' });guards is a required argument with no default. A tool set that can spend money should
never be constructible by accident, so omitting it is a type error rather than an
unlimited budget.
Full walkthrough: docs/agent-quickstart.md
| Path | What it is |
|---|---|
sdk/ |
@fatstack/x402 — provider middleware and agent client |
integrations/ai-sdk-tools/ |
@fatstack/ai-sdk-tools — a Vercel AI SDK tool set that pays per call |
contracts/ |
The USDC payment splitter and its factory, with tests and deployment gates |
facilitator/THREAT-MODEL.md |
Written verdicts on six attack classes against the payment path |
docs/ |
Provider and agent quickstarts, and the registry format |
Packages: @fatstack/x402 ·
@fatstack/ai-sdk-tools
Marketplace: fatstack.net ·
Registry: registry.json
The payment path was audited against six attack classes — replay, concurrency, cross-resource substitution, verify/settle ordering, field binding, and misconfiguration. Each has a written verdict and at least one test that proves the safe behaviour, run against a real facilitator and real settlements on Base Sepolia rather than mocks.
facilitator/THREAT-MODEL.md has the verdicts. contracts/SECURITY.md covers the splitter.
One class came back vulnerable, and it is worth reading even if you never use Fatstack, because it is a property of x402 rather than of this implementation.
An EIP-3009 authorisation covers (from, to, value, validAfter, validBefore, nonce). It
does not cover what was bought. So an authorisation signed to pay for one tool was
accepted by a different, more expensive tool from the same provider — same payee, same
asset, a valid signature, and nothing in the signed material to say otherwise.
Reproduced on Base Sepolia, settled on chain:
0xd95ea771333c8f33f346c316fad30a64585156831d09e298d501c5b1e17e5659
Fixed in @fatstack/x402 0.2.0 (current release 0.2.1, which is the same code
under the correct MIT licence metadata). The middleware now binds a payment to the resource it
was quoted for and refuses a mismatch with resource_mismatch, before the facilitator is
called — so a mismatched payment costs nothing and is never settled. Nothing a 402 quotes
to agents changed.
If you built on x402 without checking that the payment in hand was quoted for the resource in hand, check now.
An earlier draft of the splitter took the provider as a call argument. That argument is not covered by the payer's signature either, so anyone observing a pending authorisation could have submitted it with their own address and taken the provider's 98%. The provider is now fixed at construction — one splitter per provider, its address derivable in advance with CREATE2 — which removes the parameter and with it the attack.
See The finding that changed the design in contracts/SECURITY.md.
Found something? Open an issue for anything already public. For an unreported vulnerability in the payment path, please report privately through GitHub's security advisories on this repository rather than in a public issue.
The splitter is small, immutable, and has no owner, no pause, no proxy, and no upgrade
path. It is covered by 34 tests including a fork test against real USDC on Base, and its
fee arithmetic is asserted against the same
vectors.json as the TypeScript implementation, so an
on-chain split and a quoted price cannot drift apart.
No third party has reviewed it. Read it yourself before depending on it — it is about fifteen lines of code that move money and a hundred lines explaining why.
Some files here differ from their counterparts in the closed repository, because operator
detail has been removed. .redactions lists them and scripts/check-redactions.sh refuses
anything that puts that material back. Run it before pushing.
MIT. See LICENSE.