Adrian

Adrian is a runtime security tool for teams whose AI agents can use tools or act on external systems. It checks planned actions against the agent's intended task and can intervene before an unsafe action runs.

• Reasoning and activity checks: Analyse reasoning traces alongside tool calls, actions and outputs, using context across the session to identify behaviour outside the agent's remit.
• Threat detection: Detect prompt injection, unsafe tool use, credential leakage and privilege escalation, including actions that violate policy.
• Execution controls: Block risky actions, pause the agent, require human approval or sandbox an action before execution.
• Agent integrations: Connect through Python SDK support for LangChain and Anthropic, a TypeScript SDK, or a Claude Code plugin that checks tool calls and can hold them for approval in the terminal.

Adrian is free and open source under Apache 2.0. The full stack can run on a single self-hosted machine through Docker, including a dashboard and a local Gemma classifier served by Llama.cpp. This deployment supports fully offline operation without managed cloud services or telemetry leaving the host.

The bundled local classifier is tested on NVIDIA GPUs. CPU-only operation is possible, but the supplied Gemma models run slowly on real workloads.