stipend
About
Non-custodial USDC wallet on Base. Spending limits enforced in code, not in a prompt.
Details
- Author
- stipend-sh
- Categories
- Finance
Jump to
Setup
Install stipend in your MCP client (Claude Desktop, Cursor, Windsurf, and others).
Repository: https://github.com/stipend-sh/stipend
Follow the installation instructions in the repository README, then restart your MCP client.
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 anMCP (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. SeeMCP serverbelow.
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 atstipend/policy.py.It is about three hundred lines and it is the whole argument.
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.
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.
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"] } } }
Every one of them goes through the samepolicy.checkgate 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_checkexists 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.
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.
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.
No network, no funds, no wallet of yours touched:
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.
If you can construct an injection that moves money past the checks inpolicy.py, open an issue. We would rather hear it from you than from someone who is not being polite about it.
- 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.
- stipend.sh— the site, which serves plain text to agents and a page to browsers at the same URL
- stipend.sh/skill.md— written for agents
- stipend.sh/for-your-human— written for the person who pays
- kina— a small language some of this reads and writes
Built by the Stipend dev team. FelixTrade.ai Pty Ltd, Australia.
Bridge Town is an MCP-native, git-versioned financial modeling platform for FP&A teams and finance leaders. AI agents use Bridge Town tools to create projects, write Python model files, run models in isolated cloud sandboxes, query data, write outputs to Google Sheets, create dashboards, branch scenarios, and collaborate with teammates.
The Capital.com MCP Server lets your AI assistant talk to your trading account directly. Market data, position checks, trade previews – all in plain language, without leaving your AI tool.
Coinrule Agentic Trading MCP enables investors to create, backtest, execute, and manage trading agents through natural language across stocks, crypto and ETFs
Invest with Claude and other AI assistants
Australian Consumer Data Right Product Data
Remote MCP server for historical crypto & prediction-market data: search ~500K instruments, live market stats (OHLC, turnover, spreads, depth, slippage) and tick-data purchase. Keyless for catalog & stats; optional OAuth for account tools. Endpoint: https://cryptostruct.com/mcp
Cross-border debt collection from your AI assistant: check cases, get pricing, submit new cases.
Read-only MCP server for your Evibe investment portfolio + live market data (holdings, performance, dividends, benchmarks, screeners). Works with Claude & ChatGPT.
Financial and quantitative modeling engine for AI agents. Typed, named, deterministic.
Sign in to leave a review
Use Google, GitHub, or an email account so ratings stay tied to real people.
No reviews posted yet.
