ardurai/ardur

Ardur is a runtime governance and evidence layer for AI agents. Instead of trusting an agent's own account of what it did, Ardur sits between the agent and the tools or systems it touches, and records verifiable proof of each action it takes. The goal is to let a team check what an agent actually did, not just what it claims to have done in its output. The project is a Python governance runtime paired with Go components for lower level capture and Kubernetes control plane pieces, and it ships adapters for popular agent frameworks including LangChain, LangGraph, and AutoGen, along with a Claude Code plugin and hook. It also includes a personal hub service and a public evidence site built with Hugo, so evidence generated by the system can be published and reviewed. According to its own test suite, the system is built to prove three things: that its governance proxy correctly enforces rules like visibility, message integrity, and rate limits, that it resists attempts to bypass those rules through prompt injection, jailbreaking, or chained tool tricks, and that real AI models routed through it can still build working multi file software while every tool call passes through governance first. The README reports specific passing test counts for these areas as of a recent hardening round in May. The project describes itself as opening in phases, meaning some parts, like fully repeatable proof recordings and production deployment packaging, are still being finished before they are presented as ready for real use. It is aimed at teams building or deploying AI agents who need a way to audit and constrain what those agents are allowed to do, rather than individual hobbyist users.