A new npm package lets AI agents set up escrow payments on Base without holding keys
The arl-protocol-mcp package lets an agent check work and draft escrow payments, while the user's own wallet keeps the final say. It is the developer's own claim, not an audit.
The short version
- arl-protocol-mcp works with Claude Desktop, Cursor and other MCP clients, and needs Node.js 22 or later.
- Seven tools cover checking tasks, reading jobs and agent records, checking evidence, and preparing payments.
- The prepare tools return unsigned transactions; the user's own wallet must sign them.
- Every answer read from the chain names the block it came from, so anyone can re-check it.
- Agents that do not use MCP can reach the same data through an HTTP API.
Why it matters
Letting an AI agent pay a stranger is risky, and escrow is one answer: the money waits until the work is checked. The design here keeps the signing step with a human, which is where a mistake would be hardest to undo. But this is a self-published package that has not been audited, so treat it as something to examine, not to trust.
What the package does
As dev.to reported, a developer has published arl-protocol-mcp on npm, the standard package registry for JavaScript. It adds a set of tools to an AI agent so it can check finished work, read records of verified jobs, and prepare escrow payments on Base Mainnet. Base is a blockchain network.
The pitch is that the agent can do all of this without ever touching a private key. A private key is the secret that authorizes spending from a crypto wallet, so keeping it away from the agent is the point of the design.
It plugs in through MCP, the Model Context Protocol, a common way to connect outside tools to AI apps. The post says it works in Claude Desktop, Cursor and any other MCP client.
The seven tools
The post lists seven tools. Some read information and some prepare actions.
- arl_check_task validates a task against its verification policy and returns its hashes (fingerprints of the data).
- arl_list_policies shows the built-in verification policies and their hashes.
- arl_get_job reports a job's status, budget, committee and verdict at one named block.
- arl_get_agent_outcomes shows an agent's record of approved and rejected jobs, counted on chain by both evaluators.
- arl_check_evidence checks an evidence package offline.
- arl_prepare_job and arl_prepare_funding are step one and step two of setting up a paid job.
A block is one batch of records on the chain. Because each chain answer names the block it was read at, anyone can look up the same moment and compare.
How escrow works without a key
Escrow means a neutral party holds the money until a condition is met. Here the condition is that the work passes verification. That gives an agent a way to hire or pay another party without simply trusting it.
The two prepare tools build the payment steps but do not carry them out. The first drafts the transaction that creates a job. The second drafts the follow-up steps: configure, set the budget, approve and fund. Each is returned unsigned, like a form that is filled in but not yet signed.
The user's own wallet then signs. According to the post, the server never holds a key and never sends a transaction. In the author's words, that keeps the agent useful without making it a custodian, meaning someone who holds your money for you.
Setup and the HTTP route
Setup is a short block of configuration in your MCP client. It tells the client to launch the package through npx, a tool that runs an npm package without a permanent install. You need Node.js 22 or later. You can set an environment variable called ARL_RPC_URL to point the server at your own Base connection instead of the default.
Agents that do not speak MCP can use an HTTP interface called the Agent API. The post gives two sample requests, one for an agent's outcomes and one for a job's details. Both are plain read requests.
The post also invites contributions. It says good first issues on GitHub pay 150, 500 or 1,500 ARL.
What to be careful about
This is the developer's own announcement. We have not audited the code, and the post does not say how disputes are handled. A claim that no key is ever exposed is only as good as the code behind it.
Any package that touches money deserves extra suspicion. A reasonable approach is to read the code, or ask someone technical to, before connecting it to anything of value. Start with the read-only tools. If you move on to payments, use a throwaway wallet holding a tiny amount.
The unsigned-transaction design does help. Even if an agent misbehaves, nothing moves until a person approves it in a wallet. That approval step only protects you if you actually read what you are signing.
What to watch
- Whether independent developers audit the code and confirm that no keys are handled.
- How disputes are resolved, which the post does not explain.
- Whether other MCP servers adopt the same sign-it-yourself pattern for agent payments.
Try it
Before installing anything, try the read-only HTTP request from the post: curl https://arlcoin.io/api/v1/agents/1/outcomes. If you later test payments, use a throwaway wallet with a tiny balance.
Sources and links
- Reportingdev.to: the article we sourced
First reported by dev.to. Written by the SI Agent News Desk from the reporting above.Written with AI from the source above. Every number is checked against the source. Spot a mistake? Corrections.
Liked it? Say why
Add a line for other readers.
No account needed. We never show your email. Comments with links, ads or abuse are held and not shown. Your comment shows after a quick check.