Guardian
Verified Safeby FirebirdSolutions
Overview
A comprehensive AI safety system for real-time crisis detection and support, preventing AI hallucination of crisis resources and providing verified regional support.
Installation
python -m guardian_llm.cli "I'm feeling hopeless"Environment Variables
- ANTHROPIC_API_KEY
Security Notes
The project demonstrates a strong focus on AI safety, particularly against hallucination of crisis resources through a verified tool system (`tools.py`, `regions.py`). Inference is designed for on-device deployment, emphasizing user privacy by minimizing data transfer to the cloud. External API calls (Anthropic for batch data generation) correctly utilize environment variables for API keys. `subprocess.run` is used for justified, offline model export tasks, not with user-controlled input in the runtime. The inherent nature of executing model-generated tool calls is a risk vector in LLM applications, but here it's confined to a predefined set of internal, verified functions, which mitigates significant risks.
Similar Servers
bifrost
A high-performance AI gateway with a unified interface for multiple providers, offering real-time monitoring, configuration management, and comprehensive observability for AI infrastructure.
archestra
A centralized AI platform for managing, securing, and optimizing Model Context Protocol (MCP) servers and AI agents, offering features like chat interfaces, private registries, Kubernetes orchestration, dynamic tools for security, cost monitoring, and observability.
ai-counsel
Orchestrates multi-model AI deliberations to achieve deliberative consensus on complex decisions by enabling models to debate, refine positions, and leverage institutional memory.
cortexgraph
CortexGraph is an MCP server that provides short-term memory for AI assistants, featuring temporal decay, reinforcement, and automatic promotion to long-term storage.