Skip to content
ReviewRamble

The weekly brief

AI, reviewed, in your inbox.

  • New reviews of tools, models, MCP servers and repos, with the verdict up front
  • Re-checked prices and ratings when something changes
  • New analysis pieces as they are published

Models, agents, MCP servers and the tools actually worth your time.

One email a week. Unsubscribe with the link in any brief. How we use your email

OpenLIT

Self-hosted, OpenTelemetry-native tracing, evals, guardrails and GPU monitoring for LLM apps and coding agents, Apache 2.0.

Reviewed by ReviewRamble editors · Updated

View on GitHub Open source

Share this page:XLinkedInFacebookRedditWhatsAppEmail

Our rating

3.9/5

OpenTelemetry all the way down, with GPU monitoring Langfuse lacks, but a smaller community and quick-start defaults that must change before anyone else can reach it.

What is OpenLIT?

OpenLIT is two things: an SDK that auto-instruments LLM providers, agent frameworks and vector databases as standard OpenTelemetry spans, and a self-hosted platform that stores them in ClickHouse and shows traces, costs and dashboards. One openlit.init() call in Python or TypeScript starts instrumentation; there is also a Go SDK. The platform's own receiver listens for OTLP on ports 4317 and 4318, so no separate Collector is needed, and the same spans can go to Grafana, Datadog or any other OTLP backend instead.

Around tracing it adds LLM-as-a-judge and programmatic evaluations, guardrails for prompt injection and restricted topics, a versioned Prompt Hub, a Vault for API keys, a side-by-side model playground called OpenGround, a rule engine, and a GPU collector for NVIDIA, AMD and Intel cards. A CLI hooks into Claude Code, Codex, Cursor and OpenCode to trace coding-agent sessions without code changes.

The community edition is Apache 2.0 with no usage limits. Custom roles, audit logs, event alerts, notification connectors and the GPU fleet page are in a separately built Enterprise edition, licensed per organisation at prices not published. A hosted cloud is listed as coming soon. Platform version 2.2.0 shipped on 19 November 2026.

Is OpenLIT worth it?

OpenLIT's argument is that LLM telemetry should be ordinary OpenTelemetry from the first span. Its SDK emits the gen_ai.* semantic conventions, its server ingests OTLP directly, and nothing stops you sending the same data to Grafana or Datadog. Teams with an existing OpenTelemetry pipeline, or running their own models on GPUs, get more out of the box here than from most rivals: GPU utilisation, memory, power and temperature sit next to the traces they affect.

Against Langfuse, the obvious comparison, the trade is breadth against maturity. Both are free to self-host on ClickHouse and both keep audit logs and custom roles for a paid edition. Langfuse has about 36,500 GitHub stars to OpenLIT's 3,093, a hosted cloud with published per-unit prices, and deeper dataset, annotation and experiment tooling. OpenLIT has no cloud yet, publishes no Enterprise price, and its own comparison page concedes Langfuse's lead on human annotation and evaluation UI. What OpenLIT adds is GPU monitoring, guardrails, a secrets Vault and coding-agent tracing in the same install.

The catch: the quick-start defaults are not safe to expose. Docker Compose publishes ClickHouse on ports 9000 and 8123 with the password OPENLIT, the dashboard opens with user@openlit.io and openlituser, and the OTLP receiver stores spans sent without an API key. Anonymous usage metrics go to PostHog until TELEMETRY_ENABLED=false is set. In February 2026 the project published a critical advisory about a GitHub Actions workflow that exposed its own CI secrets; it concerned the build pipeline, not deployed instances.

Our take: OpenLIT is the pick for a team that already thinks in OpenTelemetry and runs its own GPUs; a team that mainly wants prompt experiments and human review will find Langfuse more finished.

Where it earns its place

  • An MLOps engineer serving Llama models on vLLM across eight GPUs

    Tie slow or expensive requests to what the hardware was doing at the time

    Run the OpenLIT GPU collector on each node and openlit.init() in the serving app; utilisation, memory and temperature land in ClickHouse next to request traces. No licence cost; the GPU fleet page itself needs Enterprise.

  • An engineering manager at a 30-developer company using Claude Code and Cursor

    See what coding agents spend in tokens and which files they touch

    Install the openlit CLI and run openlit coding install --vendor=all on each machine, pointed at a self-hosted instance. Sessions, tool calls, tokens and cost appear in the Coding Agents dashboard without changing any agent configuration.

  • A platform team already sending traces to Grafana Tempo

    Add LLM-aware views without replacing the existing pipeline

    Keep the OpenTelemetry Collector and fan out OTLP to both Tempo and OpenLIT's receiver on port 4318. OpenLIT adds cost, token and evaluation views; Tempo keeps the service-wide picture.

Before you commit

Setup
git clone and docker compose up -d starts OpenLIT, ClickHouse and NATS; a Helm chart covers Kubernetes. Before exposing it, change the ClickHouse password, NATS passwords and default dashboard login, and send spans with an API key.
Integrations
70+ auto-instrumentations across LLM providers, agent frameworks and vector databases; OTLP export to Grafana, Datadog, New Relic, SigNoz or any OTLP backend; Google, GitHub, Microsoft and Okta sign-in in the free edition.
Data and privacy
Traces stay in your ClickHouse. Anonymous product usage metrics (no prompts or span content) go to PostHog by default; TELEMETRY_ENABLED=false turns them off, with no in-app switch.
Cost at scale
No licence fee or per-trace charge for the community edition; the cost is running ClickHouse at your trace volume. Enterprise pricing is not published.
Limits
104 open issues; hosted cloud not yet launched; spans sent without an API key are stored but not evaluated; Python SDK 1.45.0 and platform 2.2.0 version separately.

What’s good

  • OpenTelemetry-native end to end, with a built-in OTLP receiver
  • GPU monitoring for NVIDIA, AMD and Intel alongside traces
  • Guardrails, Vault and coding-agent tracing in the free edition
  • Python, TypeScript and Go SDKs; 70+ auto-instrumented integrations

What’s not

  • Default ClickHouse password and dashboard login in the quick start
  • No hosted cloud and no published Enterprise pricing
  • Smaller community than Langfuse or Phoenix
  • Audit logs and custom roles are Enterprise-only

Best for

Platform teams already on OpenTelemetry, especially those serving their own models on GPUs, who want LLM tracing, evals and hardware metrics in one self-hosted stack.

OpenLIT alternatives

OpenLIT: common questions

Is OpenLIT worth it?
OpenTelemetry all the way down, with GPU monitoring Langfuse lacks, but a smaller community and quick-start defaults that must change before anyone else can reach it. We rate it 3.9 out of 5.
How much does OpenLIT cost?
Community edition free under Apache 2.0 with no usage limits. Enterprise (RBAC, audit logs, event alerts, premium connectors) priced per organisation on request; hosted cloud not yet available.
What are the downsides of OpenLIT?
Default ClickHouse password and dashboard login in the quick start; no hosted cloud and no published Enterprise pricing; smaller community than Langfuse or Phoenix; audit logs and custom roles are Enterprise-only.
Who is OpenLIT best for?
Platform teams already on OpenTelemetry, especially those serving their own models on GPUs, who want LLM tracing, evals and hardware metrics in one self-hosted stack.
How much does OpenLIT cost for a team or in production?
No licence fee or per-trace charge for the community edition; the cost is running ClickHouse at your trace volume. Enterprise pricing is not published.
Is OpenLIT safe for company data?
Traces stay in your ClickHouse. Anonymous product usage metrics (no prompts or span content) go to PostHog by default; TELEMETRY_ENABLED=false turns them off, with no in-app switch.
What are the best OpenLIT alternatives?
The closest alternatives we have reviewed are Langfuse, Arize Phoenix and OpenLLMetry.

Answers checked Nov 20, 2026.