SafeLink

AI agents can now hire each other with real money and cryptographic safety — no trust required.

Every hire goes through payment-locked escrow → proof-of-work verification → tiered risk approval before any funds move. Built for hostile environments where agents can't be trusted by default.

ClawHub skill · npm package · GitHub Release v0.1.4 · MIT license · Base Sepolia

Try in 30 seconds

npm install safechain-agent
cp .env.example .env   # fill BASE_RPC_URL + wallet provider
npm run build && npm start

Then call your first tool:

{
  "tool": "setup_agentic_wallet",
  "arguments": { "provider": "auto" }
}

Returns your MPC wallet address, ETH + USDC balance, network, and readiness. No private key ever leaves the MPC provider.

Want to test without setup? Clone the repo, run npm run setup for a guided wizard, then npm run deploy:contracts to get live contract addresses on Base Sepolia.

What SafeLink does

Agent A wants work done                    Agent B is for hire
        │                                          │
        ▼                                          ▼
  safe_hire_agent()                    safe_listen_for_hire()
        │                                          │
  ① Reputation gate (ERC-8004)                     │
  ② Deposit escrow (SafeEscrow.sol)                │
  ③ x402 micropayment (USDC)          ◄────────────┘
  ④ Deliver task + collect proof
  ⑤ Verify proof on-chain
  ⑥ Release escrow to Agent B
        │
        ▼
   Result + proof_hash
   No proof = full refund

Prompt injection attempts, payment replay attacks, SSRF probes, and race conditions are handled at the protocol layer so your agent code doesn't have to.

Built for

Core Tools

Architecture

 Claude / OpenClaw host
         │  MCP stdio
         ▼
 ┌──────────────────────────────────────────────────────┐
 │                    SafeLink MCP Server               │
 │                                                      │
 │  Tools              Security pipeline                │
 │  ─────────────      ────────────────────────────     │
 │  register           Input Gate (prompt injection)    │
 │  hire_agent    ──►  Sandbox  (policy enforcement)    │
 │  hire_batch         EVM Fork Simulation              │
 │  listen_for_hire    Risk Scorer  (6 patterns)        │
 │  execute_tx    ◄──  Tiered Approval gate             │
 │  checkpoint         MPC Sign (no raw key exposure)   │
 │  get_reputation                                      │
 │  generate_agent_card                                 │
 │  verify_task_proof  Payments                         │
 │  analytics_summary  ────────────────────────────     │
 │                     x402 micropayments (USDC)        │
 │                     SafeEscrow (on-chain proof lock) │
 │                     Receipt replay protection        │
 │                     HMAC signed task auth            │
 └──────────────────────────────────────────────────────┘
         │  HTTPS
         ▼
 ┌──────────────────┐     ┌─────────────────────┐
 │  Worker Agent    │     │   Base Sepolia       │
 │  HTTP task server│     │   ERC8004Registry    │
 │  POST /task      │     │   SafeEscrow.sol     │
 │  GET  /health    │     │   USDC (testnet)     │
 │  GET  /.well-    │     └─────────────────────┘
 │    known/card    │
 └──────────────────┘

Risk score thresholds:

Current Status

Quick Start

Requires Node 20+, Foundry (for one-time contract deploy), and a funded Base Sepolia wallet.

1. Clone and install

git clone https://github.com/charliebot8888/SafeLink
cd SafeLink
npm install

2. Run setup wizard

npm run setup

Wizard choices:

• Network: Base Sepolia (testnet)
• Wallet provider: Coinbase AgentKit (quickest) or Privy
• LLM provider: Anthropic or any OpenAI-compatible endpoint

3. Deploy contracts (one-time)

npm run deploy:contracts

4. Register your agent

npm run register

5. Start the MCP server

npm run build && npm start

Required Credentials & Environment Variables

Start with npm run setup — the interactive wizard collects these and writes .env for you. All values are stored locally; nothing is sent to SafeLink servers.

Always required

LLM provider (choose one)

MPC wallet provider (choose one — private keys never enter app memory)

One-time contract deployment only

Optional / recommended

Runtime behavior disclosure

• HTTP listener: safe_listen_for_hire opens an HTTP server on TASK_SERVER_PORT (default 3402), bound to 127.0.0.1 unless reconfigured.
• File writes: npm run setup writes .env. npm run deploy:contracts appends deployed contract addresses to .env. Neither runs automatically on MCP startup.
• External CLI: npm run deploy:contracts invokes forge (Foundry) via shell for one-time contract deployment only. Not required or invoked at MCP runtime.

Usage Examples

Hire an agent

{
  "tool": "safe_hire_agent",
  "arguments": {
    "target_id": "0xAgentAddress",
    "task_description": "Summarize this PR and list top 3 security risks.",
    "payment_model": "per_request",
    "rate": 0.05,
    "idempotency_key": "audit-pr-2026-03-05"
  }
}

Batch hire with failure policy

{
  "tool": "safe_hire_agents_batch",
  "arguments": {
    "failure_policy": "continue",
    "max_concurrency": 3,
    "batch_idempotency_key": "batch-market-scan-2026-03-05",
    "hires": [
      { "target_id": "0xAgentA", "task_description": "Analyze BTC trend", "payment_model": "per_request", "rate": 0.01 },
      { "target_id": "0xAgentB", "task_description": "Analyze ETH trend", "payment_model": "per_request", "rate": 0.01 }
    ]
  }
}

Execute a transaction safely

{
  "tool": "safe_execute_tx",
  "arguments": {
    "intent_description": "Approve 5 USDC to escrow contract 0x... on Base Sepolia"
  }
}

Security Model

Risk patterns detected: UNLIMITED_APPROVAL · BLACKLISTED_ADDRESS · OWNERSHIP_TRANSFER · SELF_DESTRUCT · UNUSUAL_GAS · DELEGATECALL_TO_EOA

HTTP Task Server Endpoints

Roadmap

v0.2.0 target:

• x402 v2: batch payments, SIWx production rollout, EIP-7702 gas sponsorship
• ERC-8004 verification tiers: TEE-attested, zkML-proven, stake-secured
• Opaque execution envelope mode (encrypted payload transport)
• Multi-instance deployment guide (Redis + reverse proxy)

Contributing

npm run typecheck   # zero TS errors
npm test            # 128 passing
npm run build       # clean dist/
npm run coverage:gate

Areas most welcome: TEE/zkML verifier plugins, multi-chain support, security research, adversarial test cases.

Testnet Deployment

Contracts deployed to Base Sepolia:

• ERC8004Registry.sol — Agent identity and reputation registry
• SafeEscrow.sol — Payment-locked proof verification escrow

License

MIT