Purser MCP Server

Purser

Give an AI agent a wallet without giving it your keys.

Purser is an ERC-4337 smart account where each agent gets its own session key with a policy: which contracts and functions it may call, how much it may spend per payment and per day, and until when. Payments inside the policy go through on their own. Anything over the limits waits for you to approve. An MCP server lets Claude, Cursor or any MCP client use the wallet, and a console shows budgets, spending and pending approvals as they happen.

Why

Agents are starting to pay for things: API credits, compute, data, subscriptions. Today that usually means one of two bad options:

• Hand the agent a private key. One prompt injection or bug away from an empty wallet.
• Approve every payment by hand. Safe, and it defeats the point of an agent.

Purser sits in between. The limits live in the contract, so they hold even if the agent, the server it runs on, or its key is compromised. The worst a leaked agent key can do is spend the rest of the current budget at allowed places, and burn some ETH on gas, until you revoke it or freeze the account, which takes one transaction. (A gas budget per key is on the roadmap.)

What's inside

How it works

flowchart LR
  Agent["AI agent<br/>(Claude, Cursor, ...)"] -- MCP tools --> Server["Purser server<br/>holds the session key only"]
  Server -- "user operation<br/>signed by the session key" --> EP["EntryPoint v0.7"]
  EP --> Wallet["PurserAccount<br/>policy checks"]
  Wallet --> Targets["Allowed contracts<br/>and recipients"]
  Owner["Owner"] -- "approve, reject, revoke, freeze" --> Console["Console"]
  Console --> Server

1. The owner grants an agent a session key with a policy: an allow-list of (contract, function) pairs, a per-transaction cap and a per-period budget for each token (ETH or ERC-20), and a validity window.
2. The agent calls MCP tools such as send_eth or call_contract. The server dry-runs the call against the policy, signs a user operation with the session key and submits it.
3. The account checks the signature, the window, the allow-list and the per-transaction caps during validation, so a disallowed operation fails before it costs anything. Budgets need the block time, which ERC-4337 forbids in validation, so they are enforced in executeUserOp.
4. Over the limit? The server turns it into an approval request. When the owner approves, the operation carries both the session signature and the owner's signature, and it runs without using the agent's budget. An approval raises the amount, never the scope: the allow-list still applies.
5. The owner can revoke a key or freeze the whole account at any time. A revoke or freeze earlier in the same bundle still stops a later agent operation.

Try it in one command

You need Docker. Nothing to clone, nothing to install:

docker run --rm -it -p 127.0.0.1:8787:8787 ghcr.io/ahmad-codex/purser:demo

The container starts a local chain, deploys the EntryPoint, the factory and a fresh account with a seeded agent, and starts the MCP server and the console. Open http://localhost:8787 and watch: ten seconds later the agent story plays in the terminal while the console updates live.

The agent pays a merchant, pays in a token, buys from a shop contract, tries a payment over the limit (the approval request appears in the console), gets approved, is refused for a recipient that is not on its allow-list, and is stopped by the kill switch. Then the wallet stays up for your own agent. Everything uses Anvil's public development keys and is gone when you press Ctrl+C.

From a clone, docker compose up runs the same stack as separate services, and docker compose run --rm demo plays the story against it.

Connect your own agent

With the demo running, point any MCP client at http://localhost:8787/mcp:

claude mcp add --transport http purser http://localhost:8787/mcp

Then ask Claude: "Check the wallet status, then pay the merchant 0.01 ETH for API credits." Try "pay the merchant 0.3 ETH" to see an approval request show up in the console. docs/mcp-clients.md covers Claude Desktop, Cursor and stdio.

MCP tools

Each payment tool returns one of: done (with the transaction hash and the budget left), waiting for the owner (with the request id), or refused (with a reason the model can act on).

The policy

Granting a key again replaces its whole policy; old permissions do not leak into the new one.

Contracts

Built on the eth-infinitism reference EntryPoint v0.7 and OpenZeppelin 5. The factory deploys through the deterministic CREATE2 proxy, so it has the same address on every chain with the canonical EntryPoint.

Tests. 41 unit tests through a real EntryPoint, fuzz tests that check random payment and time sequences against a model of the budget, and invariant tests (about 16,000 random calls per run) for three properties: an agent never spends more than its budget per period without approval, a revoked key never moves funds again, and the policy never changes behind the owner's back. Slither runs in CI. See docs/security.md for the threat model.

Deployments. See docs/deployments.md.

Repository layout

contracts/        Foundry project: PurserAccount, factory, tests, deploy and demo scripts
packages/sdk/     TypeScript client on viem: build, sign and send user operations, read state
packages/server/  MCP server (HTTP and stdio), approvals API, serves the console
apps/console/     React console for the owner
scripts/          local-up.sh, deploy-testnet.sh, demo.ts, demo-entrypoint.sh
Dockerfile.demo   all-in-one demo image (chain, contracts, server, console)
docs/             architecture, security, MCP clients, deployments

Development

npm install
cd contracts && forge test          # unit, fuzz and invariant tests
cd .. && npm test                   # TypeScript unit tests
./scripts/local-up.sh               # Anvil + deploy + demo account, writes .env.local
npm run build -w @purser/console
npm run server                      # console and MCP on http://localhost:8787
npm run demo                        # the end-to-end story; CI runs it too

No Foundry installed? contracts/forge.sh forge test runs it from the official Docker image, and scripts/local-up.sh falls back to Docker on its own.

To deploy to Base Sepolia or Sepolia, see docs/deployments.md: one script, and a faucet drip of under 0.01 ETH covers everything.

Status

Early, version 0.1.2. Not audited. Use testnets, and do not put real funds in an account you have not reviewed yourself. What is planned next is in ROADMAP.md, and issues labelled good first issue are a good place to start contributing.

Contributing

See CONTRIBUTING.md. Security reports go through private advisories, not issues.

License

MIT