wandb

This MCP server lets you query and analyze your Weights & Biases data using natural language via the Model Context Protocol. It exposes a set of tools that can fetch runs, metrics, projects, traces, and generate reports, enabling you to ask questions like: “Show runs with loss < 0.1,” “What were the top-performing runs by accuracy in project X?” or “Create a performance report for last month.” The available tools include: query_wandb_tool to query runs and metrics, query_weave_traces_tool to analyze traces and evaluations, count_weave_traces_tool to count traces and gather storage metrics, create_wandb_report_tool to programmatically produce W&B reports, query_wandb_entity_projects to list projects for an entity, and query_wandb_support_bot to access W&B documentation. When using the hosted server at mcp.withwandb.com, you can connect via the provided endpoints and supply your WANDB_API_KEY for authentication. For local setups, you’ll run the uvx-based installation locally and point your client to the local MCP endpoint, typically with a WANDB_BASE_URL override if you’re connecting to a self-hosted W&B instance.

Prerequisites:

• A host with Python and uv (the local installer used here relies on uv/uvx)
• Git access to the wandb-mcp-server repository
• A W&B API key (WANDB_API_KEY) for authentication

Option 2: Local installation (recommended for self-hosted setups)

1. Install the uvx installer if you don’t have it already. This typically comes with the uv toolchain used by Claude Desktop and other clients.
2. Install the wandb MCP server locally using the provided command:

# Install the wandb MCP server locally
uvx --from git+https://github.com/wandb/wandb-mcp-server wandb_mcp_server

3. Set your WANDB_API_KEY as an environment variable:

export WANDB_API_KEY=<your-api-key>

4. Run the MCP server locally (the installer will set up the server entrypoint). If your installer returns a local URL, note it for your MCP client configuration.
5. Configure your MCP client to point at the local server endpoint. If you’re behind a proxy or need to expose locally, consider ngrok or equivalent tunneling tools.

Option 1: Hosted server (no installation needed)

• Use the hosted server at https://mcp.withwandb.com/mcp and provide WANDB_API_KEY in your client configuration when prompted.

# Example usage in your client (OpenAI/Claude/etc.) should point to the hosted server URL:
https://mcp.withwandb.com/mcp

Tips and caveats:

• If you are connecting to a self-hosted or On-Prem W&B instance, you may need to set WANDB_BASE_URL to your instance URL (default is api.wandb.ai).
• When using local installations, ensure the WANDB_API_KEY is kept secret and not embedded in code or logs.
• The hosted server option provides zero-configuration usage, while the local path requires environment variable configuration and potentially a tunneling solution to expose a local endpoint.
• Claude Desktop users can configure their mcpServers with the uvx-based command shown in the README; ensure the full path to uv/uvx is used if Claude Desktop requires it.
• If you encounter authentication or rate-limiting issues, verify your API key permissions and ensure your account has access to the specific projects/runs you query.