What is OpenTelemetry Collector?
The OpenTelemetry Collector is a service that sits between your applications and wherever telemetry ends up. Applications send it traces, metrics and logs, usually over OTLP, OpenTelemetry's own protocol; the Collector batches them, drops or rewrites what you tell it to, and forwards copies to one or more backends. Everything is configured in a YAML pipeline of receivers (inputs), processors (filtering, sampling, redaction), exporters (outputs) and connectors (which join pipelines).
For agent teams this is where a trace that crosses a planner, three tool workers and two model providers comes back together, and where the question of where end-to-end latency went gets answered. LLM spans follow OpenTelemetry's GenAI semantic conventions, a shared set of names such as gen_ai.request.model and gen_ai.usage.input_tokens, so a Collector can send the same spans to Langfuse or Arize Phoenix for LLM debugging and to Grafana or Datadog for the rest of the stack.
The core repository holds the framework and a small set of components; most receivers, processors and exporters live in opentelemetry-collector-contrib. Both are CNCF projects under Apache 2.0. Five official distributions are published, including a core build, a contrib build and a Kubernetes build, and the Collector Builder (ocb) compiles a custom one with only the components you need.
Is OpenTelemetry Collector worth it?
The Collector is what keeps an agent system's tracing portable. Instrument once with OpenTelemetry, point every service at a Collector, and the decision of whether traces go to Langfuse, Phoenix, Datadog or all three becomes a YAML change rather than a code change. Nothing else in observability has its reach: the core repository has about 7,700 GitHub stars and contrib about 5,000, releases arrive every few weeks, and version 0.165.0 shipped on 6 November 2026.
It also does jobs an LLM backend cannot. A redaction or transform processor can strip customer data from prompt attributes before they leave your network, which matters because the GenAI conventions put full input and output messages in span attributes and warn that they may contain personal data. Sampling cuts the bill: keeping every trace of a chatty agent is expensive, and tail sampling can keep only the slow or failed ones. Langfuse's own docs show a Collector receiving OTLP over gRPC and forwarding over HTTP, because Langfuse does not accept gRPC yet.
The GenAI layer is still moving. None of the GenAI conventions has reached Stable, they moved to their own repository in 2026, and attributes have already been renamed (gen_ai.system became gen_ai.provider.name). Backends lag the spec: Langfuse still maps the deprecated gen_ai.prompt and gen_ai.completion. Pinning instrumentation versions and checking what your backend reads is part of the work. Filtering needs care too, since Langfuse needs the root span of a trace and dropping it leaves an incomplete trace.
The version number is honest about the rest. The binary is still 0.x, stability varies component by component, and breaking changes appear in release notes; 0.165.0 removed four deprecated HTTP settings. Contrib carries most of the security risk: 2026 advisories include an authentication bypass in the Azure auth extension and a GitHub receiver that ignored its configured authentication. A team with one agent service and one backend gains little from it, since the docs say sending directly is fine for small setups; the case for it starts at the second service or the second destination.
Where it earns its place
A platform team at a fintech running a support agent across six services
Trace a slow answer end to end and keep account numbers out of the trace store
Each service exports OTLP to a Collector gateway; a redaction processor masks account patterns in gen_ai.input.messages before export to a self-hosted Langfuse. The Collector itself costs only the pods it runs on.
An SRE at a retailer paying Datadog by ingested span
Keep LLM debugging detail without paying for every agent step
Send all spans to Phoenix or Langfuse for the AI team and only tail-sampled slow or failed traces to Datadog. Tail sampling needs every span of a trace on the same Collector, so a load-balancing tier comes first.
A CTO at a 15-person start-up choosing an LLM observability vendor
Avoid being locked into the first tracing tool
Instrument with OpenTelemetry and route through a Collector from the start; trialling a second backend means adding one exporter and running both side by side for a month, with no SDK swap.
Before you commit
- Setup
- Run the otelcol-contrib Docker image or binary with a YAML config listing receivers, processors and exporters; Langfuse publishes a ready sample config for forwarding to it. Kubernetes users have a dedicated distribution, and ocb builds a trimmed binary.
- Integrations
- OTLP over gRPC and HTTP in; exporters for most observability backends via contrib. Langfuse accepts OTLP over HTTP only at /api/public/otel; Phoenix and OpenLLMetry emit and read OpenTelemetry spans.
- Data and privacy
- Self-hosted; it stores nothing beyond in-memory buffers unless configured, and sends data only to the exporters you list. GenAI spans can carry full prompts and outputs, so redaction belongs here. Apache 2.0. The core repository lists two security advisories, contrib several in 2026.
- Cost at scale
- No licence or paid edition; cost is compute for agent or gateway Collectors plus the backends' ingest prices. Sizing figures for LLM workloads are not published.
- Limits
- 0.x releases with breaking changes; GenAI conventions at Development status; tail sampling needs a load-balancing layer; dropping a root span breaks traces in Langfuse. About 690 open issues and pull requests in core and 1,060 in contrib.
What’s good
- Vendor-neutral: switch or add backends without touching application code
- Redaction, filtering and tail sampling before data leaves your network
- Bridges protocols, for example gRPC in and HTTP out to Langfuse
- Custom builds with ocb keep the binary small
What’s not
- Still 0.x, with breaking changes in release notes
- GenAI conventions not yet Stable and still being renamed
- Contrib components vary in maturity and carry most advisories
- Another service to run, size and monitor
Best for
Agent teams with several services or more than one telemetry backend who want LLM traces they can redact, sample and reroute without changing code.
OpenTelemetry Collector alternatives
OpenTelemetry Collector: common questions
- Is OpenTelemetry Collector worth it?
- The standard place to collect, clean and route agent traces once more than one service or backend is involved, and an extra moving part before then. We rate it 4.4 out of 5.
- How much does OpenTelemetry Collector cost?
- Free and open source under Apache 2.0, with no paid edition from the project. Costs are the machines it runs on and the backends it sends to; commercial distributions come from observability vendors.
- What are the downsides of OpenTelemetry Collector?
- Still 0.x, with breaking changes in release notes; genAI conventions not yet Stable and still being renamed; contrib components vary in maturity and carry most advisories; another service to run, size and monitor.
- Who is OpenTelemetry Collector best for?
- Agent teams with several services or more than one telemetry backend who want LLM traces they can redact, sample and reroute without changing code.
- How much does OpenTelemetry Collector cost for a team or in production?
- No licence or paid edition; cost is compute for agent or gateway Collectors plus the backends' ingest prices. Sizing figures for LLM workloads are not published.
- Is OpenTelemetry Collector safe for company data?
- Self-hosted; it stores nothing beyond in-memory buffers unless configured, and sends data only to the exporters you list. GenAI spans can carry full prompts and outputs, so redaction belongs here. Apache 2.0. The core repository lists two security advisories, contrib several in 2026.
- What are the best OpenTelemetry Collector alternatives?
- The closest alternatives we have reviewed are OpenLLMetry.
Answers checked Nov 9, 2026.
Make OpenTelemetry Collector? Show this rating on your site
