</>macrostackBrowse all
Migration guide · LLM Observability & Evaluation

The 4 best LangSmith alternatives

LangSmith is LangChain's hosted platform for tracing, evaluating and monitoring LLM applications. It records each step of a chain or agent run, lets you build evaluation datasets and score outputs against them, and tracks token cost and latency per request — the tooling that turns 'the model said something odd' into something you can actually debug.

90
Bottom line

Strong product, and the pricing is shaped around trace volume rather than value delivered. If your traces contain customer data or your agents are chatty, a self-hosted tool pays for itself quickly.

Jump to the full comparison →

The cost

There is a free developer tier with a monthly trace allowance. Paid plans are seat-based with usage pricing on traces beyond the included volume, and enterprise plans (including self-hosted deployment) are quote-only. The cost that surprises people is trace volume rather than seats: a chatty agent generates many spans per user action, so the bill tracks how verbose your application is. Verified against LangChain's published pricing on 2026-07-30 — confirm current figures, as this pricing has changed repeatedly.

Why people consider an alternative

Usually cost shape or data residency. Trace-based billing scales with how talkative your agents are, not with how much value you get, so an application that loops can produce a bill out of all proportion to its traffic. Traces are also the most sensitive telemetry you have — they contain the actual prompts and completions, which for many teams means full customer data — and sending them to a hosted platform is a decision some organisations simply cannot make. A third reason is coupling: the tooling is built by, and around, LangChain.

When LangSmith is still the right call

If you are already building on LangChain, LangSmith is the path of least resistance and the integration genuinely is tighter than anything else. It is also the right call for teams with no appetite to run a database and a collector for their observability stack, and for anyone who needs an evaluation workflow working this week rather than assembled over a sprint.

The verdict — at a glance

AlternativeLicenseSelf-hostPricingSovereignty
Langfuse★MIT Expat (core); ee/ directories under a separate enterprise licenceYesFree to self-host — Docker Compose or Kubernetes, costs are your own infrastructure. A managed cloud with a free tier and paid plans exists if you would rather not run it.90
OpenLLMetryApache-2.0YesFree — Apache-2.0 instrumentation libraries. You pay only whatever backend you already send telemetry to; a hosted Traceloop backend is available separately.92
HeliconeApache-2.0YesFree and self-hostable under Apache-2.0. Hosted service has a free tier of roughly 10,000 requests per month, with paid plans starting around $79/month.88
Phoenix (Arize)Elastic License 2.0 (source-available, not OSI open source)YesFree to self-host under ELv2 for your own internal use. Arize's commercial platform is priced separately.62
90
Macrostack's top pick

Langfuse

The self-hosted default for LLM tracing — MIT core, framework-agnostic.

Which one fits your situation?

There is no single best answer — it depends on what you are actually trying to do.

Just tell me the best LangSmith alternative

Langfuse90

Our top pick — the highest Sovereignty Score at 90, and the one we would choose ourselves.

Langfuse vs OpenLLMetry →

I have no budget — is there a genuinely free one?

OpenLLMetry92

Open source under Apache-2.0. Free — Apache-2.0 instrumentation libraries. You pay only whatever backend you already send telemetry to; a hosted Traceloop backend is available separately.

OpenLLMetry vs Langfuse →

I want to self-host it and own the data

Helicone88

Self-hostable, so your data stays on your own server. Free and self-hostable under Apache-2.0. Hosted service has a free tier of roughly 10,000 requests per month, with paid plans starting around $79/month.

Helicone vs Langfuse →

Every alternative, compared

#1★ TOP PICK

Langfuse

The self-hosted default for LLM tracing — MIT core, framework-agnostic.

90
OPEN SOURCEMIT Expat (core); ee/ directories under a separate enterprise licenceSELF-HOSTLOCAL-FIRST

Langfuse is the most widely adopted open LLM observability platform: tracing, evaluation, prompt management and cost tracking, deployable with Docker on your own infrastructure so traces containing customer data never leave your network. It is deliberately framework-agnostic — LangChain, LlamaIndex, the provider SDKs directly, or OpenTelemetry — which makes it the natural exit from a LangChain-shaped tool. At 32k stars with commits landing daily it is the healthiest project in this category. Licensing needs precision: the core is MIT Expat, while the ee/ directories carry a separate enterprise licence.

Strengths

  • +Self-hosted: prompts and completions stay inside your network
  • +Framework-agnostic — not tied to any one orchestration library
  • +Tracing, evals, prompt management and cost tracking in one tool
  • +32k stars and daily commits — the most active project in this category

Trade-offs

  • −Self-hosting means running Postgres, ClickHouse and the collector yourself
  • −The MIT core stops at the ee/ directories; SSO and some enterprise features sit behind a separate licence
  • −Evaluation tooling is less polished than LangSmith's if you live inside LangChain
  • −Trace storage grows fast — retention policy is your problem to plan
Free to self-host — Docker Compose or Kubernetes, costs are your own infrastructure. A managed cloud with a free tier and paid plans exists if you would rather not run it.
WebsiteSourceDocs
#2

OpenLLMetry

Apache-2.0 OpenTelemetry instrumentation — send traces anywhere you already send them.

92
OPEN SOURCEApache-2.0SELF-HOSTLOCAL-FIRST

OpenLLMetry takes the opposite approach to a platform: it is a set of Apache-2.0 OpenTelemetry extensions that emit LLM traces in a standard format, so they flow into whatever you already run — Grafana, Datadog, Honeycomb, Jaeger or Traceloop's own hosted backend. If you already have observability infrastructure, this adds LLM visibility to it rather than asking you to adopt a second, parallel system. It is also the strongest hedge against lock-in: standard telemetry moves between backends by configuration.

Strengths

  • +Apache-2.0 throughout — no enterprise-directory carve-out
  • +Standard OpenTelemetry: traces go to a backend you already operate
  • +The best protection against lock-in — switching backends is configuration
  • +Reuses your existing observability stack instead of adding a parallel one

Trade-offs

  • −Instrumentation, not a platform — no evaluation or prompt-management UI
  • −You need a trace backend already, or you are choosing one anyway
  • −Generic backends render LLM traces less legibly than a purpose-built UI
  • −Smaller project (about 7k stars) than Langfuse
Free — Apache-2.0 instrumentation libraries. You pay only whatever backend you already send telemetry to; a hosted Traceloop backend is available separately.
WebsiteSourceDocs
#3

Helicone

Gateway and observability in one — one integration instead of two.

88
OPEN SOURCEApache-2.0SELF-HOSTLOCAL-FIRST

Helicone is an Apache-2.0 AI gateway with logging, cost tracking and caching built into the routing layer, which makes it a genuine two-birds option: point your client at Helicone and you get both the gateway and the observability without instrumenting your application code at all. That proxy-based approach is the practical distinction — nothing to add to your codebase, at the cost of only seeing what passes through the proxy.

Strengths

  • +Apache-2.0, fully self-hostable
  • +Proxy-based — no code instrumentation needed to start
  • +Gateway and observability in one component
  • +Real free tier on the hosted version for evaluating it

Trade-offs

  • −Only sees traffic through the proxy — no visibility into internal agent steps
  • −Weaker evaluation tooling than Langfuse or LangSmith
  • −Smaller project (about 6k stars)
  • −Adds a proxy hop to every model call
Free and self-hostable under Apache-2.0. Hosted service has a free tier of roughly 10,000 requests per month, with paid plans starting around $79/month.
WebsiteSourceDocs
#4

Phoenix (Arize)

Strong evaluation tooling — but read the licence before you build on it.

62
SOURCE-AVAILABLEElastic License 2.0 (source-available, not OSI open source)SELF-HOSTLOCAL-FIRST

Phoenix is Arize's tracing and evaluation tool, and its evaluation workflow is among the best available: run experiments, score outputs against datasets, and inspect where a chain went wrong. It runs locally in a notebook or as a self-hosted service. The critical detail most listicles omit is the licence — Phoenix is Elastic License 2.0, which is source-available rather than open source. You may use and modify it, but you may not offer it to third parties as a hosted or managed service, and you may not circumvent its licence-key functionality.

Strengths

  • +Excellent evaluation and experiment tooling — the strongest of this group
  • +Runs locally, including straight from a notebook
  • +Self-hostable, so traces stay in your infrastructure
  • +Actively developed, with commits landing daily

Trade-offs

  • −Elastic License 2.0 — source-available, NOT open source; you cannot offer it as a managed service
  • −The licence restriction is a real constraint if your product is itself a platform
  • −Score of 62 reflects the licensing limitation, not the engineering quality
  • −Smaller community than Langfuse
Free to self-host under ELv2 for your own internal use. Arize's commercial platform is priced separately.
WebsiteSourceDocs

Questions people ask

▸ What is the best open-source alternative to LangSmith?

Langfuse for most teams — MIT at the core, self-hostable, framework-agnostic, and the most actively developed project in the category at 32k stars. If you already run observability infrastructure, OpenLLMetry (Apache-2.0) is arguably the better architecture: standard OpenTelemetry traces flowing into the backend you already have, rather than a second parallel system to operate.

▸ Why does Phoenix score lower when its evaluation tooling is the best here?

Because the Sovereignty Score measures control, not quality — and Phoenix is Elastic License 2.0, which is source-available rather than open source. You can run and modify it internally, but you may not provide it to third parties as a hosted or managed service. For most internal uses that restriction never bites, and Phoenix may well be the best tool for your evaluation work. If your own product is a platform, read ELv2 carefully first. The score is a transparency lens, not a verdict.

▸ Do I need LLM observability separately from Datadog or Grafana?

Not necessarily, and this is where OpenLLMetry earns its place. LLM traces are just traces with extra structure — prompts, completions, token counts, tool calls — so OpenTelemetry instrumentation sends them into whatever you already run. A purpose-built tool renders them more legibly and adds evaluation workflows, but if you already operate Grafana or Datadog, adding LLM visibility to it is often better than running a parallel stack. Both Datadog and New Relic have been adding LLM views of their own.

▸ What actually drives the LangSmith bill?

Trace volume, not seats. Every step of a chain or agent run is a span, so a verbose agent that loops a few times can emit dozens of spans for a single user action. That means the bill tracks how talkative your application is rather than how many customers it has — which is why teams running agentic workloads are the ones who most often move to something self-hosted.

▸ Can I keep LangChain and drop LangSmith?

Yes — that is the common path and it is well trodden. Langfuse has first-class LangChain integration, and OpenLLMetry instruments LangChain through standard OpenTelemetry. Neither requires you to leave the framework. The coupling people assume exists between LangChain and LangSmith is commercial, not technical.

Compare them head-to-head

Langfuse vs OpenLLMetry→Langfuse vs Helicone→Langfuse vs Phoenix (Arize)→OpenLLMetry vs Helicone→OpenLLMetry vs Phoenix (Arize)→Helicone vs Phoenix (Arize)→

Related comparisons

Alternatives to OpenAI API (ChatGPT)8 optionsAlternatives to Notion4 optionsAlternatives to Datadog4 optionsAlternatives to 1Password4 optionsAlternatives to Google Analytics4 optionsAlternatives to Slack3 options

Entry last verified 2026-07-30. Licenses and pricing change — spotted something out of date? That's a correction we want.

The Macrostack brief

New swaps, worth your inbox.

A short, occasional email when we add a high-intent alternative or ship a new head-to-head. No spam, no selling your address — unsubscribe in one click.