📌Overview

Purpose: To provide a Model Context Protocol (MCP) server for querying Weights & Biases Weave traces and generating reports.

Overview: The Weights & Biases MCP Server enables users to access Weave traces and produce insightful reports by allowing interaction through a designated MCP client. It streamlines the process of data querying and visualization, serving as a bridge between complex analytics and user-friendly outputs.

Key Features:

Query Weave Traces Tool: Offers advanced filtering, sorting, and pagination for querying Weave traces, allowing users to retrieve complete trace data or just metadata, addressing limitations of LLM context windows.
Count Weave Traces Tool: Efficiently counts Weave traces based on specified filters, providing both total and root trace counts to inform users about project scope prior to detailed querying.
Create W&B Report Tool: Facilitates the creation of detailed W&B Reports with markdown text and visualizations, allowing for a documented analysis of findings that can be easily shared.

how many openai.chat traces in the wandb-applied-ai-team/mcp-tests weave project? plot the most recent 5 traces over time and save to a report


## Installation

```bash
git clone https://github.com/wandb/mcp-server.git
cd mcp-server && uv venv && source .venv/bin/activate
uv pip install -e .

Configuration

Create a .env file in the root directory with your Weights & Biases API key:

WANDB_API_KEY=your_api_key_here

Running the Server

Run the server using:

uv run src/mcp_server/server.py

Client Setup

Claude Desktop

Example configuration:

"mcpServers": {
    "weights_and_biases": {
        "command": "uv",
        "args": [
            "--directory",
            "/ABSOLUTE/PATH/TO/PROJECT",
            "run",
            "src/mcp_server/server.py"
        ]
    }
}

TODOs

Add W&B Models data
Convert to run with npx
Make more configurable: specify wandb URL
Improve reports plots prompt for consistent visualizations
Explore authentication solutions

Troubleshooting

Error: spawn uv ENOENT

If you see the error:

Error: spawn uv ENOENT

This means the uv package manager cannot be found.

Fix:

Install uv using the official script:

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

If the error persists, create a symlink:

sudo ln -s ~/.local/bin/uv /usr/local/bin/uv