Skip to content
MCP server

Guardian MCP

Policy gate that decides ALLOW, ASK or DENY on every agent tool call

Runs locally
Install
Platforms
CLI, API
Runs
Locally
Version
0.5.2
Listing updated
4 days ago

Similar MCP servers

About Guardian MCP

Guardian MCP is an MCP server for security. It runs locally on your machine and is started by your AI client.

Guardian sits between an agent and the tools it wants to use. The model's output is treated as a proposal: a call executes only when Guardian returns ALLOW, waits under a request ID when the answer is ASK, and is refused with a reason code on DENY. Approved ASK requests are claimed afterwards and run once.

A proxy mode wraps any stdio MCP server you already use, such as WhatsApp or Stripe, so the client still sees the upstream tools unchanged while every call is decided first. The upstream server never receives the Guardian credentials. If Guardian cannot be reached, nothing is forwarded and the result is a tool error.

A --pin-file option binds the SHA-256 of a file argument into the approval, so a file that changes after approval is caught. The package needs a Guardian control-plane URL, an agent credential and a connection id.

Install

Run in a terminal

claude mcp add guardian-mcp --env GUARDIAN_BASE_URL='<GUARDIAN_BASE_URL>' --env GUARDIAN_AGENT_CREDENTIAL='<GUARDIAN_AGENT_CREDENTIAL>' -- npx -y @odingard/guardian-mcp

Add to claude_desktop_config.json

{
  "mcpServers": {
    "guardian-mcp": {
      "command": "npx",
      "args": [
        "-y",
        "@odingard/guardian-mcp"
      ],
      "env": {
        "GUARDIAN_BASE_URL": "<GUARDIAN_BASE_URL>",
        "GUARDIAN_AGENT_CREDENTIAL": "<GUARDIAN_AGENT_CREDENTIAL>"
      }
    }
  }
}

Add to ~/.cursor/mcp.json

{
  "mcpServers": {
    "guardian-mcp": {
      "command": "npx",
      "args": [
        "-y",
        "@odingard/guardian-mcp"
      ],
      "env": {
        "GUARDIAN_BASE_URL": "<GUARDIAN_BASE_URL>",
        "GUARDIAN_AGENT_CREDENTIAL": "<GUARDIAN_AGENT_CREDENTIAL>"
      }
    }
  }
}

Add to .vscode/mcp.json

{
  "servers": {
    "guardian-mcp": {
      "type": "stdio",
      "command": "npx",
      "args": [
        "-y",
        "@odingard/guardian-mcp"
      ],
      "env": {
        "GUARDIAN_BASE_URL": "<GUARDIAN_BASE_URL>",
        "GUARDIAN_AGENT_CREDENTIAL": "<GUARDIAN_AGENT_CREDENTIAL>"
      }
    }
  }
}

Add to ~/.codeium/windsurf/mcp_config.json

{
  "mcpServers": {
    "guardian-mcp": {
      "command": "npx",
      "args": [
        "-y",
        "@odingard/guardian-mcp"
      ],
      "env": {
        "GUARDIAN_BASE_URL": "<GUARDIAN_BASE_URL>",
        "GUARDIAN_AGENT_CREDENTIAL": "<GUARDIAN_AGENT_CREDENTIAL>"
      }
    }
  }
}

Add to your client's MCP settings (most use this shape)

{
  "mcpServers": {
    "guardian-mcp": {
      "command": "npx",
      "args": [
        "-y",
        "@odingard/guardian-mcp"
      ],
      "env": {
        "GUARDIAN_BASE_URL": "<GUARDIAN_BASE_URL>",
        "GUARDIAN_AGENT_CREDENTIAL": "<GUARDIAN_AGENT_CREDENTIAL>"
      }
    }
  }
}

Using several servers? Build one config file with the MCP config generator.

Configuration

  • GUARDIAN_BASE_URLGuardian API base URL, e.g. https://guardian-api.odingard.comRequired
  • GUARDIAN_AGENT_CREDENTIALAgent credential (gak_…) from the Guardian consoleRequired
  • GUARDIAN_CONNECTION_IDTool connection id (pcn_…); required for proxy mode
  • GUARDIAN_ADAPTER_IDOptional adapter override

Tools this server provides (3)

Taken from the project documentation; the running server may differ.

guardian_invoke_tool
Proposes a tool call; ALLOW executes and returns a receipt, ASK returns a request ID held for approval, DENY refuses with a reason code.
guardian_claim
Claims an approved ASK by request ID, executes it and returns the receipt.
guardian_receipt
Fetches the machine-readable evidence receipts for a request.

What it does

ALLOW, ASK, DENY
Every call gets a deterministic decision before anything runs.
Human approval hold
ASK holds the exact call until a person approves, then it is forwarded once.
Proxy for existing servers
Puts the gate in front of any stdio MCP server without changing its tools.
Evidence receipts
Machine-readable receipts record what happened to each request.
File pinning
--pin-file hashes a file argument so edits after approval are detected.
Fail closed
Unreachable service or denial means the call is not forwarded.

Good for

  • Require human sign-off before an agent sends messages or payments
  • Put an audit trail on tools your agent already uses
  • Block risky tool calls by policy rather than by prompt
  • Check what really happened after an uncertain tool outcome

Guardian MCP FAQ

What does ASK do?

The call is held under a request ID; after a human approves it, guardian_claim forwards exactly that call once.

Does the upstream server see the agent credential?

No. In proxy mode the upstream server never receives the GUARDIAN settings.

What if Guardian is down?

Nothing is forwarded; the result is a tool error saying the call can be tried again later.

Which settings are required?

The control-plane base URL, an agent credential and a default connection id.

How do I add Guardian MCP to Claude Code?

Run this command in your terminal: claude mcp add guardian-mcp --env GUARDIAN_BASE_URL='<GUARDIAN_BASE_URL>' --env GUARDIAN_AGENT_CREDENTIAL='<GUARDIAN_AGENT_CREDENTIAL>' -- npx -y @odingard/guardian-mcp

What are the best alternatives to Guardian MCP?

The closest alternatives to Guardian MCP are Protect MCP, MCP Hangar, HOL Guard, mcpgawk and Provared MCP executor.

Where this comes from

  • Description: written from the project's README and the project's website.
  • Install steps: from the official MCP Registry entry.

How AIPoweredBot checks servers