@surfinguard/mcp-server

Surfinguard MCP server — AI agent security checks via the Model Context Protocol. Works with Claude Code, Cursor, VS Code, and any MCP-compatible client.

Install

Run directly with npx (no install needed):

npx @surfinguard/mcp-server

Or install globally:

npm install -g @surfinguard/mcp-server

Configuration

Claude Code

Add to your Claude Code MCP config:

{
  "mcpServers": {
    "surfinguard": {
      "command": "npx",
      "args": ["@surfinguard/mcp-server"]
    }
  }
}

Cursor / VS Code

Add to your MCP settings:

{
  "surfinguard": {
    "command": "npx",
    "args": ["@surfinguard/mcp-server"]
  }
}

Available Tools

The server exposes 19 tools for checking AI agent actions:

| Tool | Description | |------|-------------| | check | Universal security check for any action type | | check_url | URL phishing & impersonation detection (14 threats) | | check_command | Shell command risks — destructive ops, reverse shells (20 threats) | | check_text | Prompt injection detection (12 threats) | | check_file_read | Sensitive file exfiltration detection (10 threats) | | check_file_write | File write threats — config overwrite, persistence (12 threats) | | check_api_call | API security — SSRF, credential forwarding (10 threats) | | check_query | SQL injection & data exfiltration (7 threats) | | check_code | Code analysis — reverse shells, crypto mining (8 threats) | | check_message | Message threats — impersonation, credential sharing (6 threats) | | check_transaction | Financial transaction anomalies (6 threats) | | check_auth | Authentication risks — MFA disable, privilege escalation (6 threats) | | check_git | Git operation risks — force push, CI/CD modification (6 threats) | | check_ui_action | UI automation risks — destructive clicks, form corruption (6 threats) | | check_infra | Infrastructure risks — container escape, firewall changes (12 threats) | | check_agent_comm | Agent-to-agent communication risks (4 threats) | | check_data_pipeline | Data pipeline risks — model poisoning (5 threats) | | check_document | Document manipulation risks (4 threats) | | check_iot | IoT command risks — smart lock, vehicle injection (4 threats) |

Total: 18 analyzers, 152 threat patterns

How It Works

The MCP server runs the Surfinguard heuristic engine locally (no network calls). Each action is scored against 5 risk primitives: DESTRUCTION, EXFILTRATION, ESCALATION, PERSISTENCE, MANIPULATION.

Results include a risk level (SAFE / CAUTION / DANGER), composite score (0-10), matched threat IDs, and per-primitive breakdowns.

License

MIT