LLM OBSERVABILITY / Traceloop
OpenTelemetry MCP Server
Queries LLM application traces in Jaeger, Grafana Tempo or Traceloop, with tools that aggregate token usage by model and find slow, token-heavy or failing traces.
“An MCP (Model Context Protocol) server that connects AI assistants to OpenTelemetry trace backends (Jaeger, Tempo, Traceloop), with specialized support for LLM observability through OpenLLMetry semantic conventions.”
01 / THE REASONING
Why this made the selection.
- Published in Traceloop's own repository; pyproject.toml at this commit names opentelemetry-mcp with traceloop.com authors and license Apache-2.0, matching the LICENSE file.
- The README's tools table includes get_llm_usage, list_llm_models, get_llm_model_stats, get_llm_expensive_traces and get_llm_slow_traces, which compare models on the token counts and latency an application already records.
02 / THE REVIEW RECORD
What we actually inspected.
Source review has boundaries.
A clear record is more useful than a “safe” badge.
Material inspected
- README.md at 92ff4caf5302e3779c2050a3857b240dfc16324b
- LICENSE at the same commit
- pyproject.toml at the same commit
Our findings
- The README's HTTP mode (--transport http) serves at http://localhost:8000/sse by default; stdio is the mode for Claude Desktop.
- Configuration precedence is CLI arguments, then environment variables (BACKEND_TYPE, BACKEND_URL, BACKEND_API_KEY), then defaults. MAX_TRACES_PER_QUERY defaults to 100.
- pyproject.toml requires Python >=3.11 and declares the opentelemetry-mcp console script.
Not established by this review
- pipx or uvx launch
- Any query against Jaeger, Tempo or Traceloop
- The accuracy of token aggregation
The review applies to the material and revision named here. A newer upstream release can change its behavior.
03 / PUT IT TO WORK
Connect a server deliberately.
DOCUMENTED COMMAND
pipx run opentelemetry-mcp --backend jaeger --url http://localhost:16686 Copying does not execute this command. It may retrieve a newer version than the reviewed source.
- The README runs the server without installing it, as above, or with uvx. It was not run.
- Set BACKEND_TYPE and BACKEND_URL for your trace backend; Traceloop also needs BACKEND_API_KEY. Do not commit the key.
Before you start
- pipx or uv
- Python 3.11 or later
- A Jaeger, Grafana Tempo or Traceloop trace backend
Documented configuration example
Use the client and authentication context described in the setup guidance. Configuration formats vary between hosts.
{
"mcpServers": {
"opentelemetry-mcp": {
"command": "pipx",
"args": [
"run",
"opentelemetry-mcp"
],
"env": {
"BACKEND_TYPE": "jaeger",
"BACKEND_URL": "http://localhost:16686"
}
}
}
} 04 / FOLLOW THE EVIDENCE
The source trail.
Our notes are separate from the original resource.
Check upstream before adopting a new version.
- OpenTelemetry MCP README ↗Checked
https://github.com/traceloop/opentelemetry-mcp-server/blob/92ff4caf5302e3779c2050a3857b240dfc16324b/README.md
Supports: Supported backends, Tools table, Transports, Configuration
- OpenTelemetry MCP LICENSE ↗Checked
https://github.com/traceloop/opentelemetry-mcp-server/blob/92ff4caf5302e3779c2050a3857b240dfc16324b/LICENSE
Supports: Apache License 2.0
https://github.com/traceloop/opentelemetry-mcp-server/blob/92ff4caf5302e3779c2050a3857b240dfc16324b/pyproject.toml
Supports: Package opentelemetry-mcp, Authors at traceloop.com, License Apache-2.0