MCP Hangar v1.6 -- OTel Semantic Conventions, Transport Tracing, and the Egress-Policy Plane

v1.5 opened the front door. v1.6 is about seeing through it — and about the last mile of enforcement. Two themes: Hangar’s telemetry now speaks the OpenTelemetry GenAI/MCP semantic conventions end to end, and the L7 MCPEgressPolicy enforcement plane is shipped across all three components (core 1.6.0, operator v0.14.0, Helm charts).

Upgrade is drop-in — pip install --upgrade mcp-hangar, or pull ghcr.io/mcp-hangar/mcp-hangar:1.6.0. One breaking change for telemetry consumers is called out below and in the upgrade guide.

Telemetry that any OTel backend already understands

Before v1.6, Hangar emitted its own mcp.tool.* and mcp.cost.* span attributes and a tool.invoke.{tool} span. Correct, but Hangar-specific — a GenAI-aware backend or the OTel Collector needed custom mapping to make sense of it.

v1.6 aligns tool-invocation telemetry to the OTel GenAI/MCP semantic conventions:

• The tool name is gen_ai.tool.name; token usage is gen_ai.usage.input_tokens / output_tokens.
• The span is execute_tool {tool} with gen_ai.operation.name and mcp.method.name.
• Hangar’s governance namespaces that have no semconv equivalent — mcp.enforcement.*, mcp.risk.*, mcp.audit.*, mcp.session.id — are unchanged.

This is breaking if you query or alert on Hangar’s traces by attribute. Rename mcp.tool.name to gen_ai.tool.name (and the cost-token attributes) in your saved queries. The upgrade guide has the full list.

Tracing the upstream call boundary

The outgoing call to an upstream MCP server is now a proper SpanKind.CLIENT span, and W3C trace context propagates across the boundary on both transports — HTTP headers as before, and now the stdio transport via the MCP _meta field. So a distributed trace stays connected from the agent, through Hangar, to a stdio subprocess server. OTEL_TRACES_SAMPLER / OTEL_TRACES_SAMPLER_ARG are honored, and a failed tool call now marks its span ERROR — so failures are findable in Jaeger/Tempo instead of looking like successful traces.

There are new per-server transport metrics too — mcp_hangar_messages_sent_total, messages_received_total, and a mcp_hangar_message_size_bytes histogram — and two telemetry-health alerts (MCPHangarTelemetryExportFailing, MCPHangarDiscoveryValidationFailing) that page you when the pipeline itself breaks. Three never-emitted metrics were removed; drop any dashboard panel that referenced them.

The egress-policy plane, end to end

MCPEgressPolicy — the L7 half of Hangar’s enforcement plane — is now shipped across the stack. Declare, per MCP server or server group, a deny-by-default allow-list of upstreams (matched against the same digest/issuer primitives you already pin), plus deterministic tool-call and argument rules and a network backstop. The operator (v0.14.0) compiles the policy and pushes it to core, which enforces every tool call at the invocation chokepoint before it touches the wire. Audit-by-default, so you can watch before you block.

The Helm chart now installs the MCPEgressPolicy CRD and grants the operator the RBAC to reconcile it, so the feature works from a plain helm install. See the Egress Policy guide to get started.

Get it

pip install --upgrade mcp-hangar            # or: ghcr.io/mcp-hangar/mcp-hangar:1.6.0
helm repo update && helm upgrade mcp-hangar mcp-hangar/mcp-hangar

As always, everything is MIT and self-hosted — no SaaS.