Adrian

Why we included this project

Prompt injection and poisoned tool calls are the kind of failure that only shows up after an agent has already acted, and static analysis rarely catches them. Adrian sits at runtime instead, watching both what an agent does and the reasoning behind it, and it can block or hold a risky action before it lands. You get Python and TypeScript SDKs plus a native Claude Code plugin, so wrapping an existing agent takes a couple of lines, and terminal tool calls can be secured without touching the code. The engine is AARM-aligned and runs fully self-hosted with a local Gemma model, which is the option to pick when data sovereignty matters. Because it judges each action against the agent's intended remit rather than a list of known attack patterns, it is a deployable guardrail for production agents, not another scanner.