MLflow 3.15.0 ships MCP Registry, multimodal LLM judges, faster artifacts

MLflow, the open-source MLOps platform widely used for experiment tracking and model lifecycle management, released v3.15.0 with a batch of major features. The headline addition is an MCP Registry — a centralized catalog for registering, versioning, and sharing Model Context Protocol servers. MCP (Model Context Protocol) is an emerging standard that lets AI assistants like Claude Code connect to external tools and data sources. The registry supports semantic-versioned configs, promotable aliases, tags, auto-discovered tools, and even generates connection instructions for Claude Code and .mcp.json files. Everything is manageable via UI, REST API, or Python, making it easier for teams to standardize and share AI tool integrations.

Performance and evaluation also got significant upgrades. A new proxy-less artifact transfer mode uses presigned URLs to let large uploads talk directly to cloud storage like S3, bypassing the tracking server entirely — reducing server load and avoiding timeouts on big files, with automatic fallback to proxied transfer. For AI evaluation, the LLM judge system (make_judge()) now supports multi-modal attachments: trace spans can capture images, fetch them as base64 data URLs, and pass them to vision-capable models from Anthropic, Gemini, or OpenAI-compatible endpoints via litellm. This enables automated evaluation of screenshots, UI flows, and other visual outputs. The built-in MLflow Assistant also gained multi-provider support (Claude Code, Codex, OpenAI-compatible), live token/cost display, and simpler one-prompt setup via mlflow agent setup. Rounding things out: Pydantic AI 2.x autologging, per-endpoint budget policies in the AI Gateway, sharable saved table views, and a Cursor model catalog with pricing.