This is a notebook

Open in Colab

View in Github

The Weights & Biases MCP Server

The W&B MCP Server enables AI assistants like Cursor, Windsurf, Claude Code and Claude Desktop to directly query and analyze your Weights & Biases data. This gives your AI coding assistant powerful capabilities to help you understand experiments, debug issues, and generate insights from your W&B Models and W&B Weave data.

note

The wandb MCP server documentation on GitHub contains a fuller, in-depth guide on installation, available tools and troubleshooting.

Why use the W&B MCP Server?

When building AI applications, you often need to:

• Compare hyperparameters across experiment runs
• Analyze Weave evaluation traces to debug LLM applications
• Create visualizations of training metrics
• Generate reports summarizing experiment results

The MCP server lets your AI assistant do all of this directly, without you having to manually copy data or write analysis code from scratch.

Available tools

The server provides four main tools your AI assistant can use:

query_wandb_tool

Query W&B experiment tracking data including runs, sweeps, and metrics. Your assistant can find the best performing models and compare metrics and hyperparameters across experiments.

query_weave_traces_tool

Access Weave traces and evaluations for debugging LLM applications. Analyze latency, token usage, error rates, and trace through complex LLM workflows.

execute_sandbox_code_tool

Run Python code in secure sandboxes to perform custom analysis, create visualizations, and process data. Supports both cloud (E2B) and local (Pyodide) execution environments.

create_wandb_report_tool

Generate shareable W&B Reports with visualizations and analysis that you can share with your team.

Quick start

1. Install the uv package manager:

   curl -LsSf https://astral.sh/uv/install.sh | sh

2. Get your W&B API key from wandb.ai/authorize

3. Configure your AI assistant by adding the MCP server to its configuration file:

   {
     "mcpServers": {
       "wandb": {
         "command": "uvx",
         "args": ["wandb-mcp-server"],
         "env": {
           "WANDB_API_KEY": "your-api-key"
         }
       }
     }
   }

4. Restart your AI assistant to load the server

1-line Quickstart helpers for Cursor, Windsurf, Claude and more

The full wandb MCP server documentation on GitHub contains 1-line quickstart helpers for Cursor, Windsurf, Claude Code and Claude Desktop.

Example queries

Once configured, you can ask your AI assistant questions like:

• "What are the top 5 runs by validation accuracy in my dog-labs/pug-classification project?"
• "Show me all Weave traces where latency exceeded 2 seconds in the last hour"
• "Create a scatter plot comparing learning rate vs final loss for all runs"
• "Generate a report summarizing the performance of different model architectures"

Learn more

For detailed configuration options, sandbox setup, and troubleshooting, see the full documentation on GitHub.