Problem: you want every MCP call traced in Datadog APM with mcp.* attributes. Spans should show up in the same dashboards your other services use.

Solution: point mcpgw at your Datadog Agent’s OTLP/HTTP receiver. The Agent forwards spans to Datadog over its existing pipe — no separate ingestion path.

On the Datadog Agent — enable OTLP/HTTP

In datadog.yaml:

otlp_config:
  receiver:
    protocols:
      http:
        endpoint: 0.0.0.0:4318

Or, with environment variables (Helm chart, container):

DD_OTLP_CONFIG_RECEIVER_PROTOCOLS_HTTP_ENDPOINT=0.0.0.0:4318

Restart the Agent. Confirm the receiver is up:

curl -s -o /dev/null -w "%{http_code}\n" -X POST http://<agent>:4318/v1/traces
# Expect 415 (the receiver is up but rejects an empty body) — anything else means mis-wired

On mcpgw — point at the receiver

telemetry:
  customer:
    enabled: true
    endpoint: http://datadog-agent:4318      # bare host:port — /v1/traces is appended
    service_name: mcpgw                       # appears in Datadog APM service list
    resource_attrs:                           # optional: arbitrary OTel resource attrs
      env: production
      version: v1.0.0
      team: platform

Restart the gateway (telemetry endpoint is not hot-reloadable). Within ~10 seconds of the next request, spans appear in Datadog APM under service mcpgw.

What the spans look like

Every JSON-RPC call produces one span:

Name: mcp.tools.call (or mcp.tools.list, mcp.resources.read, etc.)
Kind: SERVER (inbound)
Attributes: mcp.method, mcp.tool.name, mcp.session.id, mcp.transport, mcp.upstream, mcp.policy.decision, mcp.payload.bytes_in/out

Outbound calls to the upstream are recorded as CLIENT child spans, so you see end-to-end latency: client → gateway → upstream → gateway → client.

See Reference: telemetry for the full attribute table and enum values.

Verifying

# Make a tool call
curl -s -X POST http://localhost:7332/mcp \
  -H "Content-Type: application/json" \
  -H "Mcp-Session-Id: dd-verify" \
  -d '{"jsonrpc":"2.0","id":1,"method":"tools/list"}'

# Then in Datadog APM:
# - service:mcpgw operation:mcp.tools.list
# - filter @mcp.session.id:dd-verify

Span should appear within 10s.

Common dashboards to build

Policy mix — pie chart of @mcp.policy.decision values. Surfaces the ratio of allow / deny / redact / rate_limit_blocked.
Top denied tools — bar chart of @mcp.tool.name filtered by @mcp.policy.decision:deny. Tells you which tool names you might be over-restricting.
Tool latency P95 — mcp.tools.call p95 grouped by @mcp.tool.name. Surfaces slow upstreams.
Sessions per tool — count(distinct @mcp.session.id) by @mcp.tool.name. Detects unusual fan-out.
Error kinds — bar chart of @mcp.error.kind. upstream_timeout and upstream_unreachable are the ones you actually want to alert on.

Pitfalls

endpoint accepts a bare host:port — mcpgw appends /v1/traces automatically. Override with the full URL only if you front the Agent with an OTLP-compatible proxy mounted at a non-default path.
service_name is also called DD_SERVICE in Datadog terms. Use whatever name you want to see in APM. Consistency matters more than the value.
Spans fail closed. If the OTLP receiver is unreachable, the gateway logs a warning every minute and continues serving traffic. Telemetry export will never block a request.
Resource attribute names must match the Datadog mapping. env, version, and service.name are auto-recognized; arbitrary keys appear as @<key> in APM filters.

Troubleshooting

Symptom	Cause	Fix
No spans in Datadog	OTLP receiver disabled	Enable `otlp_config.receiver.protocols.http` on the Agent
Spans appear under “unknown service”	`service_name` not set or empty	Set `telemetry.customer.service_name`
Datadog shows operation name only, no `mcp.*` attrs	Service name correct but custom resource attrs missing	Datadog version may need to allow custom attrs in span tags — see Datadog Agent v7.50+
5xx in mcpgw logs about telemetry	Receiver up but Agent can’t reach Datadog	Check Agent’s own status with `datadog-agent status`

Reference: telemetry — semantic conventions
Tutorial 1 — end-to-end first trace

Enable Datadog tracing