io.github.MukundaKatta/agentguard

agentguard-mcp

MCP server for @mukundakatta/agentguard. Lets Claude Desktop, Cursor, Cline, Windsurf, Zed, or any other MCP client check whether a URL is allowed under a network-egress policy before any fetch.

npx -y @mukundakatta/agentguard-mcp

Three tools:

• check_url — single URL check: returns { allowed, reason } without making any actual request.
• check_urls_batch — batch check with per-URL decisions plus a summary.
• validate_policy — sanity-check a policy spec for empty allowlists, overly broad * wildcards, and malformed host patterns.

Add to your client

Claude Desktop

Edit ~/Library/Application Support/Claude/claude_desktop_config.json (macOS) or %APPDATA%\Claude\claude_desktop_config.json (Windows):

{
  "mcpServers": {
    "agentguard": {
      "command": "npx",
      "args": ["-y", "@mukundakatta/agentguard-mcp"]
    }
  }
}

Same shape for Cursor (~/.cursor/mcp.json), Cline, Windsurf, Zed.

Tool examples

check_url:

{
  "url": "https://api.openai.com/v1/chat",
  "policy": { "allow": ["api.openai.com", "*.anthropic.com"] }
}

Returns:

{ "allowed": true, "reason": "matched_allowlist", "detail": null }

check_urls_batch:

{
  "urls": [
    "https://api.openai.com/v1/chat",
    "https://evil.example.com/leak"
  ],
  "policy": { "allow": ["api.openai.com"] }
}

Returns:

{
  "results": [
    { "url": "https://api.openai.com/v1/chat", "allowed": true, ... },
    { "url": "https://evil.example.com/leak", "allowed": false, "reason": "not_in_allowlist", ... }
  ],
  "summary": { "total": 2, "allowed_count": 1, "denied_count": 1 }
}

validate_policy:

{ "policy": { "allow": ["*", "https://api.example.com", "api.example.com/v1"] } }

Returns issues for the * wildcard, the scheme prefix, and the path suffix — common mistakes when first writing a policy.

Why a separate MCP server

@mukundakatta/agentguard is a zero-dependency JavaScript library. This MCP server makes its decision engine accessible from any MCP-aware AI assistant: ask Claude "does my agent's tool list pass this firewall?" or "which of these 50 URLs would my policy block?" and the assistant calls these tools directly.

Note: this MCP server only checks URLs — it does not actually wrap fetch or block real requests. For runtime enforcement, use @mukundakatta/agentguard directly inside your Node process.

Sibling MCP servers

Part of the agent-stack series:

• @mukundakatta/agentfit-mcp — Fit it.
• @mukundakatta/agentguard-mcp — Sandbox it. (this)
• @mukundakatta/agentsnap-mcp — Test it.
• @mukundakatta/agentvet-mcp — Vet it.
• @mukundakatta/agentcast-mcp — Validate it.

License

MIT