Heimdall — MCP proxy that exports AI agent tool calls as OTel spans to Jaeger (OTLP)

[enmanuelmag]

Hi Jaeger community,

Sharing a project I built that might be interesting to folks here: Heimdall, an OpenTelemetry proxy for the MCP protocol (used by AI assistants like Claude Desktop to call external tools).

MCP agents are opaque. When Claude calls a filesystem tool, a GitHub API, or a database server, there's no structured trace of what happened, just unstructured logs at best.

Heimdall sit transparently between the MCP client and any MCP server, intercept every JSON-RPC message, and emit a proper OTel span for each tool call.

Jaeger integration:

Since Jaeger supports OTLP natively (gRPC on :4317, HTTP on :4318), you can point Heimdall directly at it:

heimdall \
  --store sqlite://~/.heimdall/traces.db \
  --otlp http://localhost:4318/v1/traces \
  -- npx -y @modelcontextprotocol/server-filesystem /tmp

Each tool call becomes a span with:

• service.name / service.version / service.namespace via @opentelemetry/semantic-conventions
• Unix nanosecond timestamps as BIGINT
• Integer SpanKind (2 = CLIENT) and SpanStatusCode values
• Tool name, input, output, and timing in span attributes

The result is a standard Jaeger trace showing every agent action — searchable, timing-annotated, and correlatable with backend service spans if you propagate context.

NEW: Config Policies

Beyond tracing, the latest release adds per-server allow/deny rules so you can restrict what tools an agent is allowed to call. Think of it as a policy layer on top of the observability layer.

Docs: https://stack.cardor.dev/heimdall
Repo: https://github.com/enmanuelmag/heimdall-mcp

Happy to hear thoughts on the span schema or Jaeger compatibility — especially from anyone running large-scale OTLP pipelines!