MCP Security Audit for AI Agents

Your agents are wired into files, databases, email, Slack and GitHub. Paste the MCP configs from Claude, Cursor or VS Code and see every server, what it can reach, which tools have more access than the job needs, and what a prompt injection could do with all of it together.

Runs in your browser: configs and tokens never leave this page
Where to find your config files
AppFile
Claude Desktop~/Library/Application Support/Claude/claude_desktop_config.json (macOS)
%APPDATA%\Claude\claude_desktop_config.json (Windows)
Claude Code.mcp.json in the project, ~/.claude.json, and .claude/settings.json for permissions
Cursor~/.cursor/mcp.json and .cursor/mcp.json in the project
VS Code.vscode/mcp.json, or mcp in your user settings.json
Windsurf~/.codeium/windsurf/mcp_config.json
Cline / Roocline_mcp_settings.json / .roo/mcp.json
Drop config files here, or paste below.

MCP Guard: A Firewall for Agent Tool Calls

The audit shows what your agents could do. MCP Guard decides what they may do. It sits between the agent and each MCP server on your machine, reads every tool call, and stops the risky ones before the server sees them.

Install in one command

It wraps every server in the config and keeps a backup. Restart the app afterwards.

npx -y phishclean-mcp-guard@1.0.0 wrap ~/.cursor/mcp.json
npx -y phishclean-mcp-guard@1.0.0 wrap .mcp.json
npx -y phishclean-mcp-guard@1.0.0 wrap "%APPDATA%\Claude\claude_desktop_config.json"

Rules for the whole team

Commit a policy to .phishclean/mcp-policy.json in the repo and every developer's guard follows it. The audit above writes a starting policy for your own servers.

{
  "alertWebhook": "https://hooks.slack.com/services/…",
  "servers": {
    "postgres": { "readOnly": true },
    "github":   { "readOnly": true, "allowWrites": ["add_issue_comment"] },
    "fetch":    { "allowHosts": ["docs.python.org", "*.mozilla.org"] }
  }
}

See what your agents did

mcp-guard report          # calls per server and tool, blocks, injection attempts
mcp-guard log --blocked   # everything that was stopped, and why

MCP Guard is free and open source (MIT), with no dependencies, and runs entirely on your machine. Remote (HTTP) servers can be guarded through mcp-remote; see the README on npm. It can't make a shell server safe: an agent with a shell can be told to do anything a shell can, so remove those or run them in a container.

Why this matters now

MCP servers turned agents from chat windows into programs with your access. Every tool a server offers is described to the model in plain text, and every result comes back as plain text, so a web page, an email or a GitHub issue can carry instructions the model may follow. Researchers have shown tool descriptions that quietly ask the agent to read SSH keys, and a GitHub issue in a public repo that made an agent copy data from the owner's private repos into a public pull request. The fix isn't a smarter model; it's giving each agent only the access its job needs, and checking the calls it makes.

Check the apps your agents build

Building with Lovable, Bolt or Cursor? App Check scans the live app for leaked keys, open databases and API routes without sign-in. For coding-agent backdoors in a repo (.claude/settings.json hooks, dev containers), use the AI repo scanner.

Last updated: