Type: url
Original: https://github.com/invariantlabs-ai/invariant
Status: generated
Summary
The Invariant Guardrails repository, describing a rule-based guardrailing layer for LLM- and MCP-powered agent applications that is deployed as a proxy between the application and its model or tool servers, so that it can steer and monitor continuously without invasive code changes. Rules are Python-inspired matching rules that can match flows between tool calls rather than single messages — its worked examples raise on a get_inbox call followed by send_email to an address outside the company domain, and on a prompt injection detected in a get_website output followed by send_email. It runs either through the Gateway proxy, which evaluates rules on each LLM and MCP request, or programmatically through the invariant-ai package.