@arbiterlabs/mcp

Judgment tools for any MCP-capable agent — Claude Code, Claude Desktop, OpenClaw, or your own host.

Lets a model check whether an action is safe before taking it, rather than finding out afterwards.

Setup

// claude_desktop_config.json, or .mcp.json
{
  "mcpServers": {
    "arbiter": {
      "command": "npx",
      "args": ["-y", "@arbiterlabs/mcp"],
      "env": {
        "ARBITER_URL": "https://arbiter-hs23.onrender.com",
        "ARBITER_PRIVATE_KEY": "base64-64-byte-algorand-key",
        "ARBITER_MAX_SPEND_USD": "10"
      }
    }
  }
}

Tools

| Tool | Does | Cost | |---|---|---| | judge_transaction | Is this transaction safe to sign? | $0.002 | | judge_counterparty | Is this address who I think it is, and can it receive? | $0.01 | | judge_human | What would a person say about this? | $0.25 | | retrieve_human_verdict | Collect a human answer already paid for | free | | arbiter_budget | How much has been spent this session | free |

Verdicts come back as readable text with the decision first and every finding explained, because that text is what the model reasons over.

Configuration

| Variable | Default | | |---|---|---| | ARBITER_URL | http://localhost:4021 | Deployment to call | | ARBITER_PRIVATE_KEY | — | Base64 Algorand key used to pay | | ARBITER_MAX_SPEND_USD | 25 | Lifetime cap for this process | | ARBITER_MAX_PRICE_PER_CALL | 1.00 | Per-call cap |

Paying

Calls are settled per request in USDC on Algorand over x402. There is no API key and no account — payment is the authentication.

Your payer account needs USDC and must be opted in to the USDC asset. On Algorand a transfer to an account that has not opted in is rejected outright, so an un-opted-in payer cannot pay at all. Fees are sponsored by the facilitator, so the payer does not need ALGO.

| Route | Price | |---|---| | Transaction verdict | $0.002 | | Counterparty verdict | $0.01 | | Human judgment | $0.25 |

Full documentation: https://arbiter-hs23.onrender.com/docs

Licence

MIT