Weights & Biases MCP

Weights & Biases is an experiment tracking and model observability platform, and Weave is its companion layer for tracing and evaluating LLM applications. The official W&B MCP server puts both behind a single Model Context Protocol connection, so an assistant can read your training history and your production traces without a human exporting CSVs or clicking through the web app first. The server exposes experiment tracking data, meaning projects, runs, sweeps, artifacts, and reports, through a query interface flexible enough to compare arbitrary runs and metrics. Alongside it sits access to Weave traces and evaluation results for debugging generative applications. It can also write back: an agent can assemble a report containing line plots, bar plots, and run comparisons, which persists an analysis where the rest of the team will actually see it rather than leaving it in a chat log. Documentation search is bundled into the same connection, so a model can look up how a W&B concept works and then query that concept in the same turn. That combination suits ML and applied AI teams who want an agent to answer questions like which sweep configuration regressed, or which evaluation cases a change broke, and then leave a durable artifact behind. Access follows the API key used, so the agent sees the entities and projects that key already covers.