Mandoline MCP Server enables AI assistants to reflect on, critique, and improve their own performance using Mandoline's evaluation framework through the Model Context Protocol. Offered as a hosted service (recommended for most users) and optionally as a local deployment for development or contribution, it integrates with popular assistants such as Claude Code, Codex, Claude Desktop, and Cursor via standard MCP configurations. The server provides core tooling for evaluation workflows, including a health check endpoint to verify connectivity, a metrics subsystem to create and manage evaluation criteria (create_metric, batch_create_metrics, get_metric, get_metrics, update_metric), and an evaluations subsystem to score prompts and responses against those metrics (create_evaluation, batch_create_evaluations, get_evaluation, get_evaluations, update_evaluation). Documentation resources (llms.txt and MCP guides) are available to help with setup, best practices, and usage. For local use, you can run Node.js 18+ locally, start the server, and configure clients to point to http://localhost:8080/mcp. The README also includes concrete client setup snippets for Claude Code, Codex, Claude Desktop, and Cursor to establish connections to the hosted or local MCP server.