MCP server

Give any MCP client the guard as a tool: it can check an action before running it, or check every call a script is about to make. One address, your existing API key, no code.

Connect your client

You need an API key from the dashboard (new workspaces get 1,000 free requests). Pick your client, paste your key if you like, and copy the snippet.

claude mcp add --transport http mcp-guard https://api.mcp-guard.ai/mcp --header "Authorization: Bearer mcpg_live_…"

Claude Code in one line

claude mcp add --transport http mcp-guard https://api.mcp-guard.ai/mcp --header "Authorization: Bearer mcpg_live_…"

This lets Claude call the guard when it thinks to. To check every tool call, whether or not the model asks, use the PreToolUse hook instead (or as well).

The three tools

Pick a tool to see what it does, its parameters and an example call.

{
  "jsonrpc": "2.0",
  "id": 1,
  "method": "tools/call",
  "params": {
    "name": "guard_check",
    "arguments": {
      "action": {
        "tool": "db.drop_table",
        "args": {
          "table": "users"
        }
      },
      "intent": "List the tables in the staging database",
      "context": "Database agent. Environment: production."
    }
  }
}

{
  "id": "grd_3f9c0a…",
  "verdict": "block",
  "p_unsafe": 0.9597,
  "reasons": [
    "destructive",
    "arguments not supported by the request",
    "scope violation"
  ],
  "scores": {
    "safe": 0.0403,
    "violation": {
      "none": 0.0234,
      "policy_violation": 0.0768,
      "scope_violation": 0.383,
      "injection": 0.0552,
      "goal_drift": 0.3826,
      "corrigibility": 0.079
    },
    "severity": {
      "expected": 2.61,
      "level": "high",
      "probabilities": [
        0.03,
        0.08,
        0.14,
        0.75
      ]
    },
    "destructive": 0.9987,
    "exfiltration": 0.0214,
    "injected": 0.0311,
    "approval_policy": {
      "auto_approve": 0.0122,
      "require_human": 0.1105,
      "reject": 0.8773
    },
    "blast_radius": {
      "expected": 1.93,
      "level": "production-mutating or external side effect",
      "probabilities": [
        0.01,
        0.05,
        0.94
      ]
    },
    "args_grounded": 0.2135
  },
  "model": "mcp-guard-deberta-v1",
  "latency_ms": 14.8,
  "credits_charged": 1
}

guard_check takes the same fields as POST /v1/guard; guard_batch takes up to 64 of them. Results have the same shape as the REST API, plus credits_charged, free_request and free_requests_remaining. What the scores mean: scores & verdicts.

Billing, limits and errors

Tool calls go through the same pipeline as POST /v1/guard: the same key, rate limit, credit balance and usage records (logged with the endpoint names mcp and mcp/batch). 1 credit per check, or one free request per tool call while the workspace has them. Failed calls are free; usage is always free.

Unknown methods return JSON-RPC error -32601; an unknown tool name or non-object arguments return -32602. Batches (a JSON array of messages) are supported and run in order. More in errors, limits & retries.

Protocol details (for client builders)

Streamable HTTP transport, JSON responses only. Protocol version 2025-06-18; clients asking for 2025-03-26 or 2024-11-05 get that version back. Supported methods: initialize, notifications/initialized, ping, tools/list, tools/call, and empty resources/list and prompts/list. A session id is returned on initialize but never required.

POST https://api.mcp-guard.ai/mcp
Authorization: Bearer mcpg_live_…
Content-Type: application/json
Accept: application/json, text/event-stream

{"jsonrpc":"2.0","id":1,"method":"initialize","params":{"protocolVersion":"2025-06-18","capabilities":{},"clientInfo":{"name":"my-app","version":"1.0"}}}

→ 200, Mcp-Session-Id: mcp_…
{"jsonrpc":"2.0","id":1,"result":{"protocolVersion":"2025-06-18","capabilities":{"tools":{}},"serverInfo":{"name":"mcp-guard","title":"MCP Guard","version":"1.0.0"},"instructions":"…"}}

Questions