08 -- Observability: Langfuse¶

Prerequisite: 01 -- HTTP Gateway You will need: Running Hangar, Langfuse instance (cloud or self-hosted) Time: 10 minutes Adds: Distributed tracing for tool invocations via Langfuse

The Problem¶

You know a tool call was slow. You don't know whether the delay was in Hangar (routing, cold start) or in the MCP server itself. You need end-to-end traces that break down each phase.

The Config¶

# config.yaml -- Recipe 08: Langfuse Tracing
mcp_servers:
  my-mcp:
    mode: remote
    endpoint: "http://localhost:8080"
    health_check_interval_s: 10
    max_consecutive_failures: 3

observability:                           # NEW: Langfuse tracing
  langfuse:                              # NEW: Langfuse configuration
    enabled: true                        # NEW: enable Langfuse adapter
    public_key: ${LANGFUSE_PUBLIC_KEY}   # NEW: from environment
    secret_key: ${LANGFUSE_SECRET_KEY}   # NEW: from environment
    host: "https://cloud.langfuse.com"   # NEW: Langfuse host

Try It¶

1. Set environment variables:

export LANGFUSE_PUBLIC_KEY="pk-lf-..."
export LANGFUSE_SECRET_KEY="sk-lf-..."

1. Start Hangar:

mcp-hangar serve --http --port 8000

1. Make a tool call:

curl -X POST http://localhost:8000/api/mcp_servers/my-mcp/start

1. Open Langfuse dashboard and find the trace. You see spans for:
2. hangar.tool_invocation -- overall call
3. hangar.cold_start -- MCP server initialization (if cold)
4. hangar.mcp_server_call -- actual MCP server communication

What Just Happened¶

The TracedMcpServerService wraps tool invocations with Langfuse spans via the LangfuseObservabilityAdapter. Each tool call creates a trace with child spans for cold start (if needed) and the actual MCP server call. Correlation IDs link Hangar traces to MCP server-side traces.

Key Config Reference¶

What's Next¶

You've set up external observability. Now try running MCP servers as local subprocesses instead of remote HTTP.

--> 09 -- Subprocess MCP servers