Weights & Biases
Weights & Biases
Connect your AI agents to Weights & Biases through its vendor-hosted MCP server.
Generic MCP server
Weights & Biases builds and hosts this MCP server. Agent Handler proxies your agent’s calls to it, so you get the same auth, tool access controls, and audit trail you get on a Merge Connector. The tool surface itself comes from Weights & Biases.
Query experiment tracking runs, sweeps, and model registry data.
Authentication
Weights & Biases authenticates with an API key. Your end user supplies the key during Magic Link, and Agent Handler stores it encrypted.
Available tools
Tools on a generic MCP server are defined by Weights & Biases, not by Merge, so the list changes when Weights & Biases changes it. Agent Handler discovers the current tools when the connection is made and re-discovers them on each resync, which means your agent always sees what the server actually exposes today.
To see the tools available right now, connect Weights & Biases in the dashboard and open its Connector page, or read Weights & Biases’s own MCP documentation.
How this differs from a Merge Connector
Merge Connectors are written and tested by Merge against the vendor’s REST API. Merge owns the tool names, the input schemas, the error messages, and the regression tests, so the surface stays stable and the tool descriptions are tuned for agents.
A generic MCP server is the vendor’s own. You get whatever the vendor ships, on the vendor’s release schedule. Agent Handler still handles credential storage, OAuth, tool policy, and audit logging identically, so the governance story is the same either way.
See Connectors overview for the full comparison.
Next: Connect a Connector