Back to Home
wandb icon

wandb-mcp-server

Verified Safe

by wandb

Overview

The server enables natural language querying and analysis of Weights & Biases data, specifically focusing on ML experiment tracking (W&B Models) and LLM/GenAI application observability (W&B Weave) through the Model Context Protocol.

Installation

Run Command
uvx wandb_mcp_server --transport http --host 0.0.0.0 --port 8080

Environment Variables

  • WANDB_API_KEY
  • MCP_SERVER_LOG_LEVEL
  • WANDB_SILENT
  • WEAVE_SILENT
  • WANDB_DEBUG
  • MCP_LOGS_WANDB_ENTITY
  • MCP_LOGS_WANDB_PROJECT
  • WANDB_BASE_URL
  • WF_TRACE_SERVER_URL
  • MCP_AUTH_DISABLED
  • SESSION_TTL_SECONDS
  • MAX_SESSIONS_PER_KEY
  • MCP_SERVER_ENABLE_HMAC_SHA256_SESSIONS
  • MCP_SERVER_SECRETS_PROVIDER
  • MCP_SERVER_SECRETS_PROJECT
  • WANDBOT_BASE_URL
  • PARSE_ARGS_AT_IMPORT
  • SESSION_PREFIX_LENGTH
  • MCP_TRACE_LIST_OPERATIONS
  • WEAVE_DISABLED

Security Notes

The server demonstrates robust security practices, particularly for multi-tenant environments. It utilizes `ContextVar` for per-request API key isolation, preventing cross-request data leakage in concurrent operations. The `create_report` tool explicitly patches the `wandb_workspaces` API client to also use `ContextVar`, addressing known singleton contamination vulnerabilities and handling markdown input carefully. The `query_wandb_tool` allows arbitrary GraphQL queries, which is a powerful but potentially risky feature; however, its usage is heavily documented with critical warnings for the LLM to manage context windows and avoid open-ended queries. Session management includes optional HMAC-SHA256 verification via a secrets resolver. Limited `subprocess.run` calls are for low-risk operations (e.g., `git rev-parse HEAD`). No direct `eval` or `os.system` for user-controlled input was found, and sensitive secrets are expected to be managed via environment variables or a secrets resolver.

Similar Servers

Stats

Interest Score53
Security Score9
Cost ClassMedium
Avg Tokens15000
Stars4
Forks0
Last Update2025-11-25

Tags

Weights & BiasesWeaveLLM ObservabilityExperiment TrackingMCP Server