agent-firewall-mcp

Input/output safety gate for AI agents, as an MCP server. The 2026 doctrine (OWASP LLM01): treat every external input the agent touches as hostile. Google measured a +32% surge in prompt-injection payloads embedded in web content over three months of 2026.

{ "mcpServers": { "agent-firewall": { "command": "npx", "args": ["-y", "agent-firewall-mcp"] } } }

Tools

• scan_content — detect prompt injection / jailbreak / obfuscation (zero-width, bidi, hidden HTML) in text or a fetched URL → allow/review/block.

• scan_secrets — detect leaked API keys/tokens/private-keys + PII (Luhn-checked cards, SSNs, emails) → findings + a redacted copy.

• check_url — URL/domain safety: heuristics + domain age (RDAP) + redirect chain → safe/suspicious/malicious.

• check_ip — IP reputation: Tor exit, ASN/org (Team Cymru), reverse DNS, datacenter, blocklist → verdict.

• check_password — is a password in a known breach? HIBP Pwned Passwords (k-anonymity; plaintext never leaves the server).

Deterministic, free, no LLM. Data: HIBP, RDAP, Tor Project, Team Cymru, DNS + curated rulesets. Backed by the live API https://agent-firewall-seven.vercel.app (override with AGENT_FIREWALL_API). Part of the Agent Tools family. MIT.