parad0x-labs/openclaw-skills

Review: REVIEW.md lists what each component does, what has been demonstrated, where it runs and what is not established (generated from evidence/claims.json).

openclaw-skills is a set of drop-in skills that let AI agents on OpenClaw, and any MCP client such as Claude Desktop, Cursor or Windsurf, pay for services, charge for their own work, and keep long sessions cheap.

Agents increasingly call paid APIs and other agents, but they have no wallet-safe way to settle a bill. These skills add one: the agent pays in USDC on Solana through your own wallet, under a hard spending limit, and both sides keep a matching receipt. No platform holds the funds.

At a glance

How it works

One paid request, end to end. x402-pay runs in the buying agent, x402-gate in the selling one.

%%{init: {'theme':'base','themeVariables':{'primaryColor':'#111111','primaryTextColor':'#f0efeb','primaryBorderColor':'#92aa7c','lineColor':'#a3a3a3','secondaryColor':'#0a0a0a','tertiaryColor':'#0a0a0a','fontFamily':'JetBrains Mono, monospace'}}}%%
sequenceDiagram
    participant A as Your agent (x402-pay)
    participant W as Your wallet
    participant S as Paid API or agent (x402-gate)
    participant C as Solana
    A->>S: Request a resource
    S-->>A: 402 Payment Required, price in USDC
    A->>W: Unsigned USDC transfer, checked against your cap
    W-->>A: Signed transaction
    A->>C: Broadcast payment
    A->>S: Retry with payment proof
    S->>C: Verify the payment
    S-->>A: Resource and receipt hash

Both sides compute the same 32-byte receipt hash with no shared state, so each keeps a matching record. Each payment also carries a 0.05% protocol fee leg in the same atomic transaction. Full walkthrough: The Web0 agent loop.

Quickstart

Pay x402 endpoints from an OpenClaw agent.

npm i @parad0x_labs/openclaw-x402-pay

import plugin, { setX402Signer } from "@parad0x_labs/openclaw-x402-pay";

// Wire your own wallet in at startup. The skill only ever hands you a serialized tx to sign.
setX402Signer({
  publicKey: myWallet.publicKey.toBase58(),
  signTransaction: async (txBase64) => myWallet.signSerialized(txBase64),
});

The agent now has a pay_x402({ url }) tool. Payments run on devnet until you set allowMainnet: true; maxAmountUsdc caps each payment (default 1.0). Config reference: skills/x402-pay.

Use the tools from Claude Desktop, Cursor or Windsurf. Add to your MCP client config:

{
  "mcpServers": {
    "parad0x": {
      "command": "npx",
      "args": ["-y", "@parad0x_labs/mcp-server"],
      "env": { "SOLANA_RPC_URL": "https://solana-rpc.publicnode.com" }
    }
  }
}

SOLANA_RPC_URL is used for read lookups. No tool submits a transaction while receipt anchoring is unavailable, so no keypair is needed. Details: skills/mcp-server.

The skills

Companion: @parad0x_labs/null-mcp resolves .null names read-only from any MCP client. The full catalog, the module contract and the stack map are in docs/SKILLS_CATALOG.md.

Status

The vault appliance

The second product in this repo is Liquefy, a Python flight recorder for agent runs, for teams that need an audit trail of what their agents did. It is independent of the skills above: they never import it, and three small skills (liquefy-openclaw, liquefy_archive, liquefy_token_guard) are its front-ends. It packs a run folder into compressed, optionally encrypted .null vaults, checks every archive by decompressing it and comparing hashes before accepting it, and keeps a SHA-256 hash-chained audit log. Vaults can be Ed25519-signed and their fingerprints anchored on Solana as 64-bit hash prefixes in an SPL Memo transaction (default cluster mainnet; pass --cluster devnet for tests), so a third party can confirm a vault existed without seeing its contents.

It ships 24 format-aware compression engines (JSON, logs, SQL, VPC flow, CloudTrail, screenshots and more) and a set of guards: policy enforcer with kill switch, context gate with replay blocking, safe run with automatic rollback, state and history guards, PII redaction and log de-noise.

git clone https://github.com/Parad0x-Labs/openclaw-skills
cd openclaw-skills
make setup
make quick DIR=~/openclaw/sessions

For developers

Every skill is a self-contained module: no imports across skills, its own version, README and SKILL.md, and its own path-filtered CI lane. Rules in skills/README.md.

Run the tests. Each TypeScript package tests on its own:

cd skills/x402-pay && npm ci && npm test      # same pattern for each package under skills/ and plugins/

The appliance runs under pytest, plus the red-team release gate (mutation, crash-recovery and stress runs):

pip install -r requirements.txt pytest
pytest tests/
PYTHONPATH=tools:api bash tests/redteam/run_all.sh

GitHub Actions on main: CI for the appliance (Linux and macOS unit tests, CLI contracts, end-to-end smoke, red-team gate, benchmark regression), Skills (TypeScript) per changed skill, and Security (CodeQL, pip-audit).

More: docs/WEB0_AGENT_LOOP.md · docs/PARADOX_STACK.md · docs/SKILLS_CATALOG.md · PUBLISH_RUNBOOK.md · THREAT_MODEL.md

Security

Money-touching skills are non-custodial: your wallet signs, the skills never hold or read a private key, and the paying side enforces a hard USDC cap before any transaction is built. The appliance fails closed: missing encryption keys stop the server, uploads are size-capped and path-sanitized, and no data is sent back to Parad0x Labs. Report vulnerabilities as described in SECURITY.md. MIT licensed, see LICENSE.

---

Parad0x Labs · parad0xlabs.com