This MCP server provides a natural language interface for MLflow tracking, enabling large language models to query experiments, inspect runs, analyze metrics, and browse the model registry without writing code. It exposes a rich set of tools to list and search experiments, discover metrics and parameters, retrieve detailed run data, filter and sort runs, and perform side-by-side comparisons. Users can access metric histories, compare parameters across runs, and identify the best-performing models. Artifacts from runs can be browsed and downloaded, while registered models and their versions, along with deployment stages, are accessible through the MCP endpoints. It supports tag-based search and offset-based pagination to handle large result sets. The server can be installed with uvx (recommended) or via pip, and run from source. It requires a running MLflow tracking server, with MLFLOW_TRACKING_URI set (environment variable or Claude Desktop config). Health checks are available to verify connectivity. Example tool names include get_experiments, get_runs, query_runs, get_best_run, compare_runs, get_run_artifacts, get_registered_models, and get_model_version.