Lazaretto

ClaudeCommunityby Lazaretto

Lazaretto checks what an AI agent is about to trust. Give it a remote MCP server URL and it reads the advertised tool names, descriptions, parameters and server instructions for tool poisoning. It looks for hidden directives, attempts to extract secrets, instructions targeting another server's tools, and invisible Unicode. Paste an existing tool list to check a local server without contacting it. Check exact npm package versions or a lockfile against published OSV / OpenSSF malicious-package advisories. Look up a SHA-256 in a daily known-bad indicator set, find remote MCP servers from the official registry with inspection results, or retrieve and verify signed Lazaretto attestations. Checks use fixed rules and published evidence. Findings identify the matching rule and affected tool or line. No match is not a guarantee of safety. Unchecked items and stale attestations are identified. No account or sign-in is required. All seven tools are read-only. They do not install packages, execute target tools, change files or move money. Lazaretto receives the inputs supplied to its tools; it does not request your chat history. Do not include credentials or personal information in those inputs.