Repository avatar
AI Tools
v0.4.10
active

x402-guard

io.github.base-buzz/x402-guard

x402 payment safety: judge a payment against your policy before signing. Read-only, never signs.

Documentation

PayStage

Your agent's wallet should not sign because a server asked it to.

PayStage sits between an agent and the x402 payment it is about to make, and answers one question before the signature exists: should this payment happen? Every tool returns pay, review, or do_not_pay, with the rule that fired and a receipt you can keep.

The gap it fills is narrow and load-bearing. x402 makes paying trivial and leaves deciding to the caller: a wallet enforces signatures, not intent. @paystage/x402-guard evaluates the challenge before the wallet signs, so a wrong payee, an over-cap amount or an unknown asset never reaches the signing step at all.

Where the money does not move

The design point is the arrow that is missing. The wallet signs only after the challenge has been judged; a refusal returns the seller's own 402 with a verdict header attached, so there is nothing to revoke afterwards.

sequenceDiagram
    autonumber
    participant A as Agent
    participant S as Seller
    participant G as PayStage guardFetch
    participant W as Wallet

    A->>S: request the paid resource
    S-->>A: 402 + payment-required challenge
    A->>G: the same request, wrapped
    G->>G: decode challenge, evaluatePayment(challenge, policy)
    alt do_not_pay, or review with onReview refuse
        G-->>A: original 402 + x-paystage-verdict
        Note over G,W: nothing is signed
    else pay, or review with onReview pay
        G->>W: x402-fetch stages the payment
        W-->>S: signed payment
        S-->>G: 200 + resource
        G-->>A: 200 + x-paystage-verdict
    end

review is reported as review either way. When onReview: "pay" lets an ambiguous payment through, the header still says review, because the evaluator's finding and the caller's instruction to tolerate it are different facts — a caller that cannot tell "clean" from "warned but paid" cannot audit itself. The rule that fires travels in x-paystage-reason-codes.

What decides

%% tight ranks: this is read at a glance, not studied
%%{init: {"flowchart": {"nodeSpacing": 26, "rankSpacing": 30}}}%%
flowchart TD
    C[402 challenge] --> E["evaluatePayment<br/>guards: agent budget, time window,<br/>x402Version 2, accepts present"]
    E --> P{any accept judged pay?}
    P -- yes --> PAY[pay, cheapest accept]
    P -- no --> R{any accept judged review?}
    R -- yes --> REV[review]
    R -- no --> DN[do_not_pay]
    REV --> O{onReview}
    O -- pay --> SIGN[x402-fetch signs, request retries]
    O -- refuse --> X[refused, 402 returned unsigned]
    PAY --> SIGN
    DN --> X

do_not_pay is a real refusal, not a warning: a payee on the known-drainer or OFAC list, an over-cap amount, a wrong network or asset, a policy that denies the payee, or a challenge that cannot be parsed. review is the narrow middle — a payee the policy has no opinion about, under a policy that chose unknown_payee: "review" over the stricter default. Both verdicts and the branch logic are in packages/preflight/src/evaluate.ts; the hook that uses them is packages/x402-guard/src/index.ts.

Try it without installing anything:

curl -s -X POST https://api.paystage.xyz/v1/evaluate_mcp_payment \
  -H 'content-type: application/json' \
  -d '{"payment_required":{…},"policy":{…}}'

The hosted endpoint serves a real verdict with no payment and no account for the first 1,000 calls per IP per day — an agent cannot fill in a signup form, so requiring one would exclude exactly the callers this is built for.

The ten tools

Nine are read-only. None of them signs, holds a key, or moves money — the agent stays the only signer, which is the entire product.

The tenth, request_testnet_funds, is the one deliberate exception and it is worth naming rather than hiding: it asks our Base Sepolia faucet for 0.5 test USDC so you can exercise the paid path before spending anything. It moves testnet funds from a wallet we hold, touches nothing you own, and cannot reach mainnet.

All ten are listed with what each answers in Tools. 语言 / Language: English · 简体中文

See it work

npm run build -w @paystage/preflight && node examples/run.mjs

Four challenges against one policy, four verdicts with reasons — no network, no keys. See examples/.

Quick start

# Option 1: run instantly
npx -y -p @paystage/mcp paystage-mcp

# Option 2: install locally, then use the package bin
npm install @paystage/mcp
npx paystage-mcp

The server waits silently for an agent app to connect. That's normal; press Ctrl+C to exit. To try the tools without an app, run npx @modelcontextprotocol/inspector npx -y -p @paystage/mcp paystage-mcp.

@paystage/mcp is the umbrella MCP server (10 tools, including evaluate_hip3_trade and evaluate_velocity). It exposes the PayStage tools over stdio and reads a local policy file. See SECURITY.md for exactly what it accesses and why.

Connect in one command

Copy the line for your client. The remote is Streamable HTTP, so nothing needs installing.

# Claude Code
claude mcp add --transport http paystage https://api.paystage.xyz/mcp

# Cursor — append to .cursor/mcp.json
{"mcpServers":{"paystage":{"url":"https://api.paystage.xyz/mcp"}}}

# Any MCP client, or just to read the tool schemas without an account
npx -y mcporter list https://api.paystage.xyz/mcp

Prefer it local? npx -y -p @paystage/mcp paystage-mcp runs the same ten tools over stdio and reads a policy file from disk.

What it does not catch

A gate that claims to stop every bad payment is the thing this project argues against. LIMITS.md is the honest boundary — settlements that already happened, payees that look legitimate and do not deliver, payments made outside the wrapper, and eight more. Read it before you trust the gate with money.

Every term this project uses — x402, facilitator, settlement, EIP-3009, HIP-3, CAIP-2 and the rest — is defined in one page.

State of x402 — issue 1: the registry counted, with its method.

Verdict vocabulary

Three states, and every project in this space picked its own three words. If you run more than one of these gates, here is the mapping — the semantics line up more closely than the names suggest.

PayStageMeaningSame intent elsewhere
paywithin policy — sign and sendDingDawg ELIGIBLE · Black_Wall GO · MandatePay authorize
reviewa human should look before payingDingDawg STEP_UP · Black_Wall HOLD · MandatePay challenge
do_not_payrefuse — a rule fired, and it is namedDingDawg DENY · Black_Wall STOP · MandatePay deny

One difference worth knowing: review is a policy instruction, not an oversight. onReview: "pay" treats it as permission, "refuse" treats it as a block — the gate never decides that for you.

How this compares with the four closest things in this lane — including what each has that we do not.

Start here if you are paying a seller

Burn five minutes before your agent pays anything — get the seller's challenge, write one policy file, ask for a verdict, and keep the receipt. curl and node, nothing to install.

Four ways in

You wantUse
the gate inside an agent's payment loop@paystage/x402-guard
the tools in Claude Desktop or Cursor@paystage/mcp
nothing installed at all — connect to a URLhttps://api.paystage.xyz/mcp (Streamable HTTP, listed in the MCP Registry as a remote server)
a verdict from your own service, hosting includedPOST https://api.paystage.xyz/v1/evaluate_mcp_payment
the evaluator as a library@paystage/preflight
the Vercel AI SDK@paystage/ai-sdk — two read-only tools, docs

Listed on

Configure

The same MCP config works for Claude Desktop (claude_desktop_config.json) and Cursor (.cursor/mcp.json).

Using npx

{
  "mcpServers": {
    "paystage": {
      "command": "npx",
      "args": ["-y", "@paystage/mcp"],
      "env": {
        "PAYSTAGE_POLICY_PATH": "/absolute/path/to/policy.json",
        "PAYSTAGE_RPC_URL": "https://mainnet.base.org"
      }
    }
  }
}

Using a local npm install

{
  "mcpServers": {
    "paystage": {
      "command": "node",
      "args": ["./node_modules/@paystage/mcp/dist/cli.js"],
      "env": {
        "PAYSTAGE_POLICY_PATH": "/absolute/path/to/policy.json",
        "PAYSTAGE_RPC_URL": "https://mainnet.base.org"
      }
    }
  }
}

The MCP server itself needs no environment variables and starts with an empty environment. It answers an MCP handshake and lists its tools with nothing configured — the policy arrives as an argument on each call, not as a file path. Verified:

env -i npx -y -p @paystage/mcp paystage-mcp   # completes `initialize`, lists ten tools

The two variables below are for the stdio CLI path, where a policy file is convenient. They are optional, and a directory that reads this section as "required configuration" will refuse to start a server that in fact needs nothing:

  • PAYSTAGE_POLICY_PATH (optional, CLI only): path to a policy.json the CLI reads at startup.
  • PAYSTAGE_RPC_URL (optional): RPC endpoint used by on-chain checks. Defaults to a public Base RPC.

Policies

Everything PayStage does is decided by one local policy file. Three complete, annotated examples are in registry/examples/ — copy one, change the addresses and amounts, and point PAYSTAGE_POLICY_PATH at it. Each is schema-exact for @paystage/policy and annotated field by field in registry/examples/README.md.

  • Reader bot — reads articles and search results; 0.05 USDC per call, unknown payees denied.
  • Trading bot — places HIP-3 orders; per-agent budgets, velocity windows and human review above 10 USDC.
  • Research agent — discovers data endpoints; unknown payees go to review, never an unattended pay.

Tools

toolwhat it answers
evaluate_mcp_paymentshould this x402 payment be signed? → pay · review · do_not_pay
list_outstanding_authswhich ERC-20 / Permit2 / EIP-3009 authorisations are still outstanding
evaluate_outstanding_authswhich of those need attention before the next payment
evaluate_hip3_tradeshould this HIP-3 / trade.xyz order be placed? → allow · review · do_not_trade
evaluate_velocityis this spend a burst, a spike, or a runaway cadence
verify_receiptis this receipt really ours — and is it about this payment, still in date
decode_payment_requiredwhat does this base64 402 header actually say
validate_policyis this policy document valid, and if not which field
request_testnet_funds0.5 Base Sepolia USDC for testing — one drip per address per 24h, with the reset time
evaluate_batchthe same question for several payments, one verdict each

Packages

Ten packages. Only @paystage/mcp exposes ten MCP tools; the rest are the evaluator, guards for other frameworks, a policy schema and a skill — hence ten packages and ten tools in the same breath.

PackageRoleLinks
@paystage/mcpumbrella MCP server — start herenpm
@paystage/x402-guardpolicy gate for x402 clients, before the wallet signsnpm
@paystage/ai-sdkthe same gate as Vercel AI SDK toolsnpm
@paystage/auth-watchoutstanding ERC-20 / Permit2 / EIP-3009 authorisationsnpm
@paystage/policypolicy schema and loadernpm
@paystage/preflightx402 payment preflight (the evaluator)npm
@paystage/skillagent skill manifestnpm
@paystage/trade-preflightHIP-3 / trade.xyz order preflightnpm
@paystage/velocityrate and spend-pattern checksnpm
@paystage/x-evmchain-agnostic EVM helpersnpm

Verify it yourself

Every claim this project makes about itself is reducible to a command. Run it before you trust anything on the website:

node apps/hosted/scripts/conformance.mjs

It checks the published surfaces against reality — manifest shape, the ten tools being listed, machine surfaces being genuinely different from each other, health and timing headers on both networks, the receipt key endpoint, that an unpaid evaluation is challenge-only, and that malformed input fails closed. Read-only: no payment, no key, no account. Exit code 1 if anything is off.

It runs daily in CI (cadence.yml) and the result history is on the site.

Supply chain

Socket.dev scores packages on capability, so a payment-safety tool will always carry flags: it reads a policy file, reads two environment variables and talks to an RPC. Every package README states exactly what it touches and why, and SECURITY.md has the full table. The umbrella @paystage/mcp is ~16 KB with no network, no filesystem and no eval of its own — the capability lives in the tools it bundles.

Links

Development

This is an npm-workspaces monorepo — package-lock.json is the lockfile and CI runs npm ci. There is no pnpm workspace and no lint script; npm run typecheck is the check that exists.

npm ci
npm run build
npm test
npm run typecheck

Contributing

PayStage is built in the open, outside contributions are welcome.

  • Add a drainer address. Open a PR adding it to KNOWN_DRAINERS.md with a public source link (incident report, explorer label, or security write-up). Unsourced addresses won't be merged.
  • Add an integration example. LangChain, AgentKit, ElizaOS and others are all welcome.
  • Report a bypass. If an unsafe payment gets past preflight, open an issue. That's the most valuable contribution there is.

Look for issues labelled good first issue to get started.

License

MIT