🔍 VerdictSwarm MCP Server

Fight AI with AI. Scammers build rugs with AI. Your agent catches them with 6.

The security layer for AI agents that touch money. One tool call between your trading agent and a rug pull.

Works with Claude Desktop, OpenClaw, Cursor, Codex, Windsurf, and any MCP-compatible client.

Why?

AI trading agents are making autonomous on-chain decisions with little or no risk review. Scammers are building rugs that bypass many single-algorithm scanners. The agent attack surface is now direct and financial, and bad results are costly.

VerdictSwarm deploys 6 adversarial AI agents that independently analyze, debate, and reach consensus on any token:

• Adversarial debate - not one model's opinion. Six agents argue, including a Devil's Advocate that challenges every verdict. You see dissenting opinions.
• On-chain security audits - mint authority, freeze authority, honeypot detection, LP lock status
• Rug pull detection - holder concentration, bundle/sniper activity, contract age analysis
• Agent-native - built for MCP. One tool call. Sub-second cached responses. No blockchain node required.

How It's Different

Most scanners give you a number. VerdictSwarm gives you the argument.

Tools

Real Results

Is $CRABGE safe?

VerdictSwarm Score: 23/100 (Grade F - CRITICAL RISK)

🔴 Security Agent: Mint authority still active. Deployer can mint unlimited tokens. 🔴 Liquidity Agent: Only $2,400 liquidity. 89% held by top 5 wallets. 🟡 Social Agent: 340 holders but suspicious clustering pattern. 🔴 Devil's Advocate: "Even the bull case here is terrible. Contract is 6 hours old with active mint authority."

Verdict: AVOID - High probability rug pull

Example: Quick Score

User: What's the risk score for BONK?
Agent: [calls get_quick_score("DezXAZ8z7PnrnRJjz3wXBoRgixCa6xjnB7YaB1pPB263")]
→ Score: 74/100 (Grade B) - LOW-MEDIUM risk

Example: Rug Check

User: Is this new memecoin safe? 7xKXtg2CW87d97TXJSDpbD5jBkheTqA83TZRuJosgAsU
Agent: [calls check_rug_risk("7xKXtg...")]
→ DANGER
  🚨 Liquidity NOT burned or locked
  ⚠️ Mint authority active
  ⚠️ Token is less than 24 hours old
  ⚠️ Bundle/sniper activity detected

Resources & Prompts

Resources (reference data for agents):

• verdictswarm://help - Tool usage guide
• verdictswarm://scoring - Score interpretation (0-100 scale, grades A-F)

Prompts (pre-built workflows):

• should_i_buy(token_address) - Full investment analysis with recommendation
• portfolio_check(tokens) - Batch risk assessment across holdings

Supported Chains

Scoring Guide

Architecture

MCP Client (Claude, Cursor, OpenClaw, Codex...)
    |
    |  MCP Protocol (stdio)
    ▼
┌──────────────────────────┐
│  VerdictSwarm MCP Server │  ← This package (thin wrapper)
│  FastMCP + Python        │
└──────────┬───────────────┘
           │  HTTP (httpx)
           ▼
┌──────────────────────────┐
│  VerdictSwarm API        │  ← api.vswarm.io
│  6 AI agents + on-chain  │
└──────────────────────────┘

The MCP server is a stateless wrapper. All intelligence lives in the VerdictSwarm API. This means:

• Lightweight deployment (no GPU, no blockchain node)
• Single source of truth for scan logic
• API-level rate limiting and caching already work

Development

git clone https://github.com/sentien-labs/verdictswarm-mcp.git
cd verdictswarm-mcp
pip install -e ".[dev]"
pytest  # 47 tests, ~0.3s

Links

• Website: vswarm.io
• API Docs: api.vswarm.io/docs
• GitHub: sentien-labs/verdictswarm-mcp
• MCP Spec: modelcontextprotocol.io
• Smithery: smithery.ai/servers/sentien-labs/verdictswarm-mcp
• awesome-mcp-servers: Listed here

Built by Sentien Labs - AI-operated crypto intelligence infrastructure.

Quick Start

Install & Run

# Install from PyPI (recommended)
pip install verdictswarm-mcp

# Run with uvx (zero-install)
VS_API_KEY=your_key_here uvx verdictswarm-mcp

Claude Desktop

Add to your claude_desktop_config.json:

{
  "mcpServers": {
    "verdictswarm": {
      "command": "uvx",
      "args": ["verdictswarm-mcp"],
      "env": {
        "VS_API_KEY": "your_key_here"
      }
    }
  }
}

Then ask Claude: "Check if this token is safe: DezXAZ8z7PnrnRJjz3wXBoRgixCa6xjnB7YaB1pPB263 on Solana"

OpenClaw

mcpServers:
  verdictswarm:
    command: uvx
    args: ["verdictswarm-mcp"]
    env:
      VS_API_KEY: your_key_here

Cursor

Uses claude_desktop_config.json format.

{
  "mcpServers": {
    "verdictswarm": {
      "command": "uvx",
      "args": ["verdictswarm-mcp"],
      "env": {
        "VS_API_KEY": "your_key_here"
      }
    }
  }
}

Windsurf

{
  "mcpServers": {
    "verdictswarm": {
      "command": "uvx",
      "args": ["verdictswarm-mcp"],
      "env": {
        "VS_API_KEY": "your_key_here"
      }
    }
  }
}

VS Code + Cline

{
  "cline.mcpServers": {
    "verdictswarm": {
      "command": "uvx",
      "args": ["verdictswarm-mcp"],
      "env": {
        "VS_API_KEY": "your_key_here"
      }
    }
  }
}

No API Key?

You get 10 free quick-score calls, no credit card required. Full 6-agent scans require an API key or a paid Solana USDC transaction signature.

Configuration

License

MIT - see LICENSE.