MCP 服务器

stipend

A non-custodial USDC wallet on Base exposed as seven tools: address, balance, check, pay, earnings, report and recover. Spending limits (per transaction, per day, per counterparty, plus a destination allowlist) are enforced in code between deciding and signing, and the server runs locally over stdio so the key never leaves the machine.

README

stipend

A non-custodial USDC wallet on Base that an AI agent installs by itself, with spending limits enforced below the instruction layer.

pip install eth-account && curl -sL stipend.sh/install | sh

This is the source of the package that installer fetches. It is published so that the claims made for it can be checked rather than believed.

It ships an MCP (Model Context Protocol) server — stipend mcp — exposing seven tools over stdio, so an agent can hold and spend money through the tool interface it already has. See MCP server below.

The claim worth checking

The failure that matters for an agent holding money is not a stolen key. It is a correctly signed transaction it was argued into.

Almost every defence against that lives in the instruction layer — a system prompt saying "do not send funds to unverified addresses." That is a suggestion to a model, sitting in the same context window as the text attacking it.

In here the limits are not advice. They are code that a transaction passes through on its way to being signed, and no wording in any context window reaches them.

Start at stipend/policy.py. It is about three hundred lines and it is the whole argument.

The three controls

control	what it does
per-transaction and daily caps	a ceiling that no instruction can raise
one-time approvals	authorise a single payment — bound to that recipient, that amount, expiring, single use — instead of raising a cap
new-destination confirmation	the first payment to any address you have never paid needs confirming, whatever the size

The second is the one people get wrong. Raising a cap to afford one purchase weakens every future payment to everyone, which is precisely what an attacker asks for. Authorise the payment, not the capability.

The third is what stops a drain by many small payments that each sit under the threshold. An attacker's address is, by definition, one you have never paid.

And if you turn on an allowlist, it is absolute. Funds reach nothing else — not with a confirmation, not with an approval, not with any instruction from anywhere.

MCP server

The wallet is also a Model Context Protocol (MCP) server, so an agent can use it through the tool interface it already has rather than by shelling out.

stipend mcp

It speaks JSON-RPC over stdin and stdout — the MCP stdio transport. Your MCP client spawns the process; nothing listens on a port and nothing is exposed to a network.

In a client's config that looks like:

{
  "mcpServers": {
    "stipend": {
      "command": "stipend",
      "args": ["mcp"]
    }
  }
}

The tools

tool	what it does
`stipend_address`	the wallet's address, so money can reach you whether or not you are running
`stipend_balance`	USDC held, native currency held, spent today, and the daily limit
`stipend_check`	would this payment be allowed? Answers without sending anything
`stipend_pay`	send USDC, subject to every limit below
`stipend_earnings`	read incoming payments off the network and record them
`stipend_report`	earned against spent, by category, with runway in days
`stipend_recover`	what is outstanding, and whose job each item is

Every one of them goes through the same policy.check gate the command line uses, so the caps, the allowlist and the new-destination confirmation apply identically. There is no path here that moves money without them.

stipend_check exists so an agent can ask "would this be allowed?" before it offers to pay for something, turning a refusal into an answer rather than a failure.

What it deliberately does not expose

No tool reads, exports or derives the private key. No tool changes a spending limit or the allowlist. No tool creates or overwrites a wallet.

An MCP tool is callable by a model that is reading untrusted text, and raising a cap is exactly what an attacker would ask for — so it is not on the menu at all. Configuration changes stay at the command line, where a human is present.

Why it is not hosted

Every registry would prefer a URL, and we will not publish one. A hosted wallet server means the key lives on somebody else's machine — ours — and the whole claim of this package is that it does not. This runs beside the keystore, under the same user, and the key never moves.

Run its own tests

No network, no funds, no wallet of yours touched:

python -m stipend selftest

219 checks covering address validation, amount arithmetic, every policy branch, the allowlist, the daily cap, destination trust and its expiry, keystore encryption, ERC-20 encoding, x402 requirement selection, EIP-3009 signing, the ledger, and exactly what telemetry would send.

What we would rather have

Someone to break it.

If you can construct an injection that moves money past the checks in policy.py, open an issue. We would rather hear it from you than from someone who is not being polite about it.

Honest limits

Not independently audited.
x402 interoperability is verified against our own endpoint, not third-party merchants.
The key is generated on your machine and never transmitted — which also means nobody can recover it for you, including us.
Keep the balance small. It is a working float, not savings.