Deterministic guards and tools for AI agents

Six small tools that AI agents and developers call constantly, each available both as a plain HTTP API and as an MCP server. No LLM in the loop, and no accounts or API keys for the free tier. Same input, same output. Boring and reliable on purpose.

Five of them are guards: one check per risky action an agent takes, such as installing a package, reading untrusted text or email, following a link, sending money, or writing code. The sixth is a utility set for reading and parsing the web. All are free to run, with paid tiers on RapidAPI for higher volume.

The tools

Quick start (MCP)

Add any or all to your client config (Claude Desktop, Cursor, Claude Code, and so on):

{
  "mcpServers": {
    "package-guard":  { "command": "npx", "args": ["-y", "package-guard-mcp"] },
    "agent-firewall": { "command": "npx", "args": ["-y", "agent-firewall-mcp"] },
    "payment-guard":  { "command": "npx", "args": ["-y", "payment-guard-mcp"] },
    "email-guard":    { "command": "npx", "args": ["-y", "email-guard-mcp"] },
    "code-guard":     { "command": "npx", "args": ["-y", "@mlawsonking/code-guard-mcp"] },
    "web-tools":      { "command": "npx", "args": ["-y", "web-tools-mcp"] }
  }
}

Each one

Package Guard (package-guard-mcp)

verify_package (does it exist, else likely a hallucination or slopsquat, with suggestions), check_vulns (OSV), package_info, audit_deps, typosquat_scan. Ecosystems: npm, PyPI, Go, crates.io, RubyGems, Maven, NuGet. Data: OSV.dev, npm, PyPI. API: https://package-guard.vercel.app. Code: package-guard-mcp/ and package-guard/.

Agent Firewall (agent-firewall-mcp)

scan_content (prompt injection, jailbreak, hidden-text obfuscation), scan_secrets (secrets and PII, with a redacted copy), check_url, check_ip, check_password (HIBP, k-anonymity). Data: HIBP, RDAP, Tor, Team Cymru, DNS. API: https://agent-firewall-seven.vercel.app. Code: agent-firewall-mcp/ and agent-firewall/.

Payment Guard (payment-guard-mcp)

screen_address (address or ENS to a safe/caution/block verdict), screen_payment (x402 or merchant URL), check_sanctioned (fast OFAC), resolve_name (ENS, screened), screen_token (honeypot, rug, and tax risk via on-chain simulation). Data: OFAC SDN

... View full README on GitHub