ai-safety-mcp-server
Verified Safeby Jack0319
Overview
A centralized Model Context Protocol (MCP) server for AI Safety research, providing knowledge base, safety evaluation, mechanistic interpretability, and governance tools for research assistants and agentic systems.
Installation
docker-compose up -dEnvironment Variables
- LITELLM_API_KEY
- SAFETY_EVAL_MODEL
- KB_VECTORSTORE_URL
- KB_COLLECTION
- INTERP_MODEL_DIR
- LOG_LEVEL
Security Notes
Secrets (e.g., LITELLM_API_KEY) are managed via environment variables. The server defaults to stdio for communication (local IPC), with TCP transport planned but not yet implemented. The README provides strong warnings against exposing the server directly to the internet, explicitly recommending deployment behind an authenticated proxy and usage of VPNs or private networks. Interpretability tools load models from HuggingFace or local paths, which requires trust in the model source, a standard practice in ML development. No direct `eval()` of user input or dangerous `subprocess` calls were identified.
Similar Servers
simplenote-mcp-server
Integrates Simplenote with Claude Desktop as a memory backend or content source via the Model Context Protocol (MCP), enabling AI interaction with user notes.
dotprompts
A personal prompt management system exposed as a Model Context Protocol (MCP) server, enabling AI agents to access, create, update, and delete user-defined prompts.
Kotak_Neo_MCP_Server_With_Agentic_Application.
A server designed to host and manage AI agentic applications, likely for automating tasks or processing information. The specific functionality is unknown due to missing code.
mcp-server-context-engineering
This project provides a server-side framework for managing and integrating contextual information, potentially involving AI models and search modules, within a larger engineering platform.