Weave MCP Server + Client Linked Traces:

What is Weave MCP Server + Client Linked Traces?

This server adapts the Arize Phoenix MCP tracing example for use with Weights & Biases (wandb/Weave). It provides an MCP client and server that export OpenTelemetry traces to Weave, enabling distributed tracing across MCP client–server boundaries.

How to use Weave MCP Server + Client Linked Traces?

Configure a .env file with your OPENAI_API_KEY and PHOENIX_COLLECTOR_ENDPOINT (and PHOENIX_API_KEY if required), install dependencies via uv sync or pip install -r requirements.txt, then run uv run client.py or python client.py from the tutorials/mcp/tracing_between_mcp_client_and_server directory. The client spins up the server automatically and you can ask questions to the agent.

Key features of Weave MCP Server + Client Linked Traces

• End-to-end tracing across MCP client and server
• Automatic OpenTelemetry context propagation via openinference-instrumentation-mcp
• Exports traces to wandb/Weave
• Supports multi-language MCP components
• Connects AI models to external data sources

Use cases of Weave MCP Server + Client Linked Traces

• Debugging requests that span multiple MCP services
• Observing latency at different stages of an MCP request lifecycle
• Troubleshooting performance bottlenecks in AI agent systems
• Understanding how different MCP components affect application behavior

FAQ from Weave MCP Server + Client Linked Traces

What does Weave MCP Server + Client Linked Traces do compared to the original Phoenix example?

It adapts the Arize Phoenix MCP tracing tutorial to export traces to Weights & Biases (Weave) instead of Arizona Phoenix, while preserving the same OpenTelemetry instrumentation approach.

What runtime dependencies are required?

Python 3.x, openinference-instrumentation-mcp, and standard MCP dependencies. Install using uv sync or pip install -r requirements.txt.

Where do traces get exported?

Traces are exported to the Weave collector endpoint specified in your PHOENIX_COLLECTOR_ENDPOINT environment variable. (Note: despite the variable name, the README states the repo is adapted for wandb/Weave.)

Is there a known limitation or bug in this server?

Yes, there is a race condition that sometimes causes the tool to fail at the OpenAI call step. This bug was also present in the original Phoenix example and was not introduced by the Weave adaptation.

How is trace context transported between client and server?

Trace context is automatically propagated across the MCP client-server boundary via OpenTelemetry context injection and extraction handled by the openinference-instrumentation-mcp package.