Creativity · MCP — server

MCP Weights & Biases Server

The MCP Weights & Biases Server (W&B) uses the wandb Python SDK to list projects, read run metrics, compare sweeps, and fetch artifacts. It turns an MCP client into an experiment-tracking copilot: 'find the best hyperparameters across my last 20 runs' becomes a single chat turn.

MCP facts

Kind: server
Ecosystem: anthropic-mcp
Language: Python (wandb SDK)
Transports: stdio

Capabilities

Tools: list_runs, get_run_metrics, compare_runs, list_sweeps, fetch_artifact
Resources: wandb://project/{entity}/{project}/run/{id}
Auth: WANDB_API_KEY (read-only recommended)

Install

pipx install mcp-server-wandb

Configuration

{
  "mcpServers": {
    "wandb": {
      "command": "mcp-server-wandb",
      "env": {
        "WANDB_API_KEY": "${WANDB_API_KEY}",
        "WANDB_ENTITY": "vset-ai-lab"
      }
    }
  }
}

Frequently asked questions

Can it start new training runs?

W&B is primarily a tracker; runs are started from training code. The MCP server can kick off a sweep agent but not an arbitrary training script.

Does it support private cloud W&B deployments?

Yes — set WANDB_BASE_URL to the private deployment and use an API key from that instance.

How is it different from MLflow MCP?

W&B is SaaS-first with strong visualisation. MLflow is OSS-first with pluggable tracking stores. Both expose similar MCP shapes; choose based on your team's tracker.

Sources

Weights & Biases — wandb SDK — accessed 2026-04-20
Model Context Protocol — accessed 2026-04-20
MCP servers repo — accessed 2026-04-20