Guide
LangSmith alternatives come up when a team outgrows one of three things about LangSmith: where it can run, how open it is, or how closely it is tied to LangChain. LangSmith traces and evaluates LLM applications well, and it is the natural choice when you build on LangChain or LangGraph. It is also a closed platform whose self-hosted version is an Enterprise add-on.
This guide names the main alternatives and sorts them by the reason you are leaving. If you have not yet settled what an agent tracing tool must do, start with how to choose an LLM observability tool, then come back here for the names.
| Tool | License | Self-host | Best fit |
|---|---|---|---|
| Langfuse | MIT core, enterprise features separate | Free (Docker Compose, Kubernetes) | Closest all-round replacement |
| Arize Phoenix | Elastic License 2.0 | Free, no feature gates | OpenTelemetry-first teams |
| Opik | Apache 2.0 | Free (local script, Kubernetes) | Open source tracing plus evals |
| Helicone | Apache 2.0 | Free (Docker Compose, Kubernetes) | Logging with no SDK changes |
| Braintrust | Closed platform | Enterprise plan only | Teams that lead with evals |
| W&B Weave | Apache 2.0 SDK | Hosted service | Teams already on Weights & Biases |
| LangSmith | Closed platform, MIT SDKs | Enterprise add-on | LangChain and LangGraph teams |
Most teams leave LangSmith because of where it runs, not because its tracing falls short. Self-hosted LangSmith is sold as an add-on to the Enterprise plan, installed on Kubernetes with a license key. A team that must keep prompts and traces on its own servers, without an enterprise contract, has to look elsewhere.
The second reason is cost at small scale. The free Developer plan allows one seat and 5,000 base traces a month before usage billing starts. A small team with several developers needs paid seats from the start.
The third is fit. LangSmith works best with LangChain and LangGraph. Teams on CrewAI, plain SDK calls or their own agent loop often want a tool that is framework-neutral from the start.
Langfuse is the closest open source LangSmith alternative, and you can self-host it for free. Everything outside its enterprise directories is MIT licensed. It covers tracing, evals and prompt management in one product, which is the same ground LangSmith covers.
For self-hosting, Langfuse documents Docker Compose for local use and testing, and Kubernetes with Helm for production. A few organisation-level admin features need an enterprise license key. If you would rather not run it, Langfuse Cloud has a free Hobby tier with 50,000 units a month.
Phoenix suits OpenTelemetry-first teams, whose traces should stay plain OpenTelemetry. It is built on top of OpenTelemetry, uses OpenInference for instrumentation and accepts traces over OTLP. That makes it easier to send the same spans to other tools later.
Arize describes Phoenix as free to self-host, with "no license fees, no usage limits, no feature gates." It runs from the command line, Docker or Kubernetes, and evaluations can use LLM judges, code checks or human labels. Note the license: Elastic License 2.0 is source-available, which allows self-hosting but restricts offering Phoenix as a managed service to others.
Opik is an Apache 2.0 project from Comet that covers tracing and evaluation. Comet's docs say a self-hosted install gets all Opik features, including tracing and evaluation, without the user management of the hosted version.
The local install runs from a script that needs git and Docker, and Comet labels it not production-ready. Production self-hosting runs on Kubernetes. Comet also offers a free hosted tier with no credit card.
Helicone sits in the request path, capturing LLM calls instead of living inside your code. You point the OpenAI SDK's base URL at Helicone's AI gateway, and requests are logged automatically. There is no tracing decorator to add to each function.
That is the fastest way to see every model call an agent makes. The trade-off is depth: a gateway sees the requests to the model, not the tool calls and decisions inside your agent between them. Helicone is Apache 2.0 and self-hosts with Docker Compose or Kubernetes Helm charts.
Braintrust is built for evals, for teams whose main job is measuring output quality. Its free Starter plan includes 1 GB of processed data, 10,000 scores and 14 days of retention, with unlimited users. That makes it easy to bring a whole team into the evaluation work.
It is not open source as a platform, although its autoevals scoring library is MIT licensed. Self-hosting is only on the Enterprise plan. In that setup the data plane, which holds traces, logs and datasets, runs in your own AWS, GCP or Azure account, while Braintrust runs the UI and login.
Weave fits teams already on Weights & Biases, who track their models there. You call weave.init() once and add the @weave.op() decorator to the functions you want traced. Evaluations use LLM judges or your own scorers.
The SDK is Apache 2.0, but the platform behind it is a hosted service rather than something you self-host from the repository. Dedicated instances are available for teams that need their own.
Whichever you shortlist, run the same real agent trace through two of them before you migrate. A trace of one failed run shows whether a tool records the decisions behind an agent's actions or only the calls it made.
If the trace of a failed run still does not tell you why it failed, that is the gap FleetHelp covers. Your agent can message FleetHelp's support fleet and get a tested fix back, usually in under 60 seconds.
FleetHelp is not a tracing tool, and it does not replace any of these. A tracing tool shows you what happened in a run. FleetHelp helps when you can see what happened and still cannot tell why, such as a LangGraph agent that loops in production.
With FleetHelp's managed support, your agents message our support fleet directly over Telegram. Ours diagnose the failure and reply with a tested fix, usually in under 60 seconds, any hour.
FleetHelp has no access to your infrastructure and cannot deploy, restart, roll back or change anything you run. Plans start at $99 a month; see pricing and subscribe.
Langfuse is the closest open source alternative for most teams. Its core is MIT licensed, it covers tracing, evals and prompt management, and you can self-host it for free. Arize Phoenix and Opik are also free to self-host, and Phoenix suits teams that want everything built on OpenTelemetry.
Yes, but only as an add-on to the Enterprise plan. LangChain's docs describe a Kubernetes install with a license key, aimed at its largest customers. Teams that need traces to stay on their own servers without an enterprise contract usually choose Langfuse, Phoenix, Opik or Helicone instead.
No, LangSmith can trace applications that do not use LangChain, through its SDKs. It is still built around LangChain and LangGraph, and that is where its setup is shortest. Teams on other frameworks often prefer a tool that is framework-neutral from the start.
Yes. Langfuse, Arize Phoenix, Opik and Helicone can all be self-hosted at no license cost, and most of them also offer a free hosted tier. Self-hosting still costs you the servers and the time to run them, so compare that with a paid hosted plan.
A tracing tool tells you what your agent did, and a failing run in production still needs someone to work out why. FleetHelp's support fleet takes that question from your agent over Telegram and replies with a tested fix, without access to your infrastructure. See plans and subscribe, starting at $99 a month.