There is no single best tool. The useful question is narrower: can you change every call site, must the data stay on your infrastructure, and is your first question about cost or about quality? Those three answers eliminate most of this list for you.
Below are twelve tools with the licence you actually get, how you wire each one in, whether self-hosting gives you the whole product, and how actively each repository is developed. GitHub figures were read on 2026-07-30.
Disclosure: we build Spanlens, which is one of the twelve. It is also the youngest and smallest project here, and the table says so.
| Tool | Stars | Licence | How you install it | Self-host |
|---|---|---|---|---|
| LiteLLM | 55,038 | Custom licence, not a standard OSI template | Run the gateway, then point clients at it and describe models in YAML | Yes |
| Langfuse | 32,119 | MIT core, plus an ee/ folder under a commercial license | Wrap clients with the SDK, or send OpenTelemetry spans | Yes, Docker Compose |
| Comet Opik | 20,966 | Apache 2.0, no gated folder | SDK decorators, plus OpenTelemetry ingest | Yes, Docker or Kubernetes |
| Portkey | 12,593 | MIT for the gateway; the observability platform is commercial | Run the gateway, or use the hosted endpoint | Gateway yes, full observability no |
| Arize Phoenix | 10,802 | Elastic License 2.0, which is not an OSI-approved open-source licence | OpenTelemetry instrumentation | Yes |
| Traceloop OpenLLMetry | 7,340 | Apache 2.0 | OpenTelemetry instrumentation libraries | Yes, and it is backend-agnostic |
| Helicone | 6,015 | Apache 2.0 | Swap the baseURL, or async SDK logging | Yes, Docker |
| Laminar | 3,127 | Apache 2.0 | SDK instrumentation | Yes |
| OpenLIT | 2,658 | Apache 2.0 | OpenTelemetry instrumentation | Yes |
| Spanlens | 10 | MIT, whole repository | Swap the baseURL, or one CLI command | Yes, one Docker command |
LangSmith and Braintrust are closed source and have no public repository, so they are not in this table. Both appear in the tool-by-tool notes below.
Pick a proxy or gateway. You change one base URL and every request through it is recorded, which matters most in an existing codebase with calls scattered across services.
Look at: Spanlens, Helicone, LiteLLM, Portkey
Langfuse has 32,119 stars and the deepest set of guides, integrations, and community answers. When you hit an unusual problem, someone has probably written about it already.
Look at: Langfuse
Choose a tool built around scoring rather than one that added it later. Datasets, judges, and experiment comparison should be the main surface, not a tab.
Look at: Comet Opik, Braintrust, LangSmith
Do not adopt a second dashboard. Emit OpenTelemetry spans with LLM attributes into the backend you already pay for and already know how to query.
Look at: Traceloop OpenLLMetry, Arize Phoenix, OpenLIT
This is a gateway problem before it is an observability problem. Retries, budgets, and provider failover come first, and logging follows through callbacks.
Look at: LiteLLM, Portkey
Per-request USD next to the model version, aggregated by prompt version and by customer, is not the default in most of these tools. Check that the cost view exists before you commit.
Look at: Spanlens, Helicone, LiteLLM
Rule out anything closed source or hosted-only first, then check whether self-hosting gives you the whole product or only part of it.
Look at: Langfuse, Comet Opik, Spanlens, Laminar
Best for. You want per-request cost in the dashboard on day one and you would rather change a base URL than wrap every call site.
The catch. It is the youngest project here by a wide margin. Ten stars against Langfuse’s 32,000 is not a rounding error, and a young tool means fewer battle-tested edge cases and a smaller community to search when you get stuck.
Best for. You want the most mature open-source option with the largest community, and you are comfortable instrumenting call sites rather than routing traffic through a proxy.
The catch. The ee/ folder gates enterprise features such as SCIM, audit logs, and project-level RBAC behind a commercial license, so self-hosting does not give you everything in the repository.
Best for. Evaluation is your main question. You want scoring, datasets, and experiment tracking as the centre of the product rather than a feature bolted onto a log viewer.
The catch. Instrumentation is SDK-first, so there is no base-URL swap. Cost tracking is present but is not the organising idea the way it is in a proxy-first tool.
Best for. You specifically want a proxy and you value a codebase that has been in production for years over one that ships quickly.
The catch. Development has slowed sharply. The repository took 24 commits between 1 May and 30 July 2026, against more than 300 each for Langfuse and Opik over the same window, and the most recent tagged release is from August 2025.
Best for. You need one endpoint in front of many providers with retries, fallbacks, and budgets, and routing matters more to you than the observability surface.
The catch. It is a gateway first. Logging exists through callbacks into other backends, so the built-in observability is thin next to a purpose-built tool, and the YAML config grows quickly.
Best for. You want a gateway with guardrails and caching and you are happy to pay for the hosted dashboard rather than run the whole stack.
The catch. The open-source repository is the gateway only; the observability you would compare against these other tools sits in the paid product. The gateway repository has had no commits since 25 May 2026.
Best for. You are already an OpenTelemetry shop and you want tracing plus evaluation that sits naturally beside your existing spans.
The catch. Elastic License 2.0 restricts offering the software as a managed service, so calling it open source is inaccurate even though the code is public.
Best for. You want LLM spans in the observability backend you already run, whether that is Grafana, Datadog, or Honeycomb, rather than adopting another dashboard.
The catch. It is plumbing rather than a product. You get well-shaped spans and then you build the cost views, evaluation, and prompt management yourself.
Best for. Multi-step agents are your whole workload and you want a backend built for that shape rather than one that treats a trace as a nice-to-have.
The catch. Small project and small community, so expect to read source when something surprises you.
Best for. You run your own models and want GPU utilisation next to token cost, which most hosted-API-focused tools do not show you.
The catch. Narrower than the larger tools once you step outside self-hosted inference.
Best for. Your stack is LangChain or LangGraph and you want the tool built by the same team, with the tightest integration available.
The catch. No public repository and no self-hosting outside enterprise agreements, so your traces live somewhere you do not control.
Best for. Evaluation is a team workflow with reviewers and datasets, and you want the polished product rather than assembling one.
The catch. Closed source, no self-hosting, and it is an evaluation product first, so request-level cost observability is not the centre of it.
Stars, licence, and commit activity come from the GitHub API, read on 2026-07-30. The activity claims are counts over a fixed window rather than impressions: for the 1 May to 30 July 2026 period, Langfuse and Comet Opik each recorded more than 300 commits, while Helicone recorded 24 and the Portkey gateway recorded 26 with none after 25 May.
Licence descriptions say what you get rather than repeating the SPDX identifier, because the identifier hides what matters. Langfuse reports as non-standard because MIT sits alongside a commercially licensed ee/ folder. Arize Phoenix uses Elastic License 2.0, which restricts offering the software as a service and is not OSI-approved.
Funding rounds, acquisitions, and valuations are deliberately absent. They change often, they are hard to verify from a public API, and a wrong claim about another company is not worth making on a page like this.
Install shape and self-hosting reflect each project's own documentation at the time of writing. If something here is out of date or wrong, tell us at support@spanlens.io and we will correct it.
There is no single best one. Langfuse has the largest open-source community, Comet Opik leads on evaluation, LiteLLM wins on multi-provider routing, and Traceloop is the choice when you want spans in an observability backend you already run. The right answer depends on whether you can change call sites, whether you must self-host, and whether cost or evaluation is your first question.
Comet Opik, Helicone, Traceloop OpenLLMetry, Laminar, OpenLIT, and Spanlens ship under OSI-approved licences with no gated directory. Langfuse is MIT with an ee/ folder under a commercial licence. Arize Phoenix uses Elastic License 2.0, which is not OSI-approved. LangSmith and Braintrust are closed source.
Proxy and gateway tools do: Spanlens, Helicone, LiteLLM, and Portkey all sit in front of the provider so you change a base URL rather than wrapping call sites. Langfuse, Opik, Laminar, LangSmith, and Braintrust use SDK instrumentation. Traceloop, Phoenix, and OpenLIT go through OpenTelemetry.
You need the token counts from the provider response joined to a price table for that exact model version, because providers return dated model ids such as gpt-4o-mini-2024-07-18. Proxy-based tools compute this at the moment the response passes through. SDK-based tools compute it wherever you instrumented.
Yes for most of this list. Langfuse, Opik, Helicone, LiteLLM, Phoenix, Traceloop, Laminar, OpenLIT, and Spanlens all run on your own infrastructure. Portkey self-hosts the gateway but not the observability platform. LangSmith self-hosts on enterprise plans only, and Braintrust does not offer it.
Head-to-head pages go deeper than the notes above, feature by feature.