dash-langsmith

Automatic LangSmith tracing for Dash MCP tool calls.

Every time an AI agent calls a tool on your Dash MCP server, a run is recorded in LangSmith — latency, inputs, outputs, and errors included. No code changes to your app required.

Install

pip install dash-langsmith

Usage

Set your API key and run your app normally.

export LANGSMITH_API_KEY=your-key
export LANGSMITH_PROJECT=my-dash-app  # optional
python app.py

That's it. The dash_hooks entry point registers the tracing middleware automatically at startup.

How it works

dash-langsmith uses Dash's plugin hooks system. On startup, hooks.setup wraps app.server.wsgi_app with a lightweight WSGI middleware that intercepts POST requests to /_mcp. When a tools/call JSON-RPC message arrives, it creates a LangSmith run before dispatching to Dash and patches it with the response on the way out.

Only tools/call requests are traced. Layout, resource, and tools/list requests pass through untouched.

Optional: richer traces with @mcp_traced

For explicit per-tool control — custom names, docstrings, per-tool project routing — use the @mcp_traced decorator instead of @mcp_enabled:

from dash_langsmith import mcp_traced

@mcp_traced(name="get_inventory", expose_docstring=True, project_name="my-dash-app")
def get_inventory(category: str) -> dict:
    """Return inventory levels for a product category."""
    ...

Configuration

Requirements

• Dash ≥ 4.3.0 (MCP support)
• langsmith ≥ 0.1.0