LangSmith Alternatives: 8 Picks by the Reason You're Leaving
November 6, 2026
The best LangSmith alternative depends on what's pushing you out. If it's the per-seat bill, look at Langfuse or Braintrust, whose plans don't charge by headcount. If you need to self-host without an Enterprise contract, Langfuse, Arize Phoenix, Opik and MLflow all run on your own servers for free. If you want instrumentation that no vendor owns, start with OpenLLMetry or Pydantic Logfire, which emit standard OpenTelemetry. If prompts are the real job, PromptLayer is built around them.
Here's the short version:
- Closest all-round swap: Langfuse (tracing, evals, prompts, free self-host)
- Lightest self-host: Arize Phoenix (one
phoenix serveprocess) - Permissive licence, full platform: Opik (Apache-2.0, backend included)
- Already on MLflow: MLflow's own tracing and evals
- Pure OpenTelemetry, any backend: OpenLLMetry
- Whole-app observability for Python teams: Pydantic Logfire
- Eval-first product teams: Braintrust
- Prompt editing by non-engineers: PromptLayer
If you've already narrowed it to Langfuse, our Langfuse vs LangSmith head-to-head goes deeper. This post starts from the other side: you're on LangSmith today, and something about it no longer fits.
Why teams leave LangSmith
LangSmith is a good product. It has grown from a tracing tool into a platform: LangChain's pricing page now lists observability and evaluation, Deployment, Fleet (no-code agents), Engine (automated issue fixing), Sandboxes and an LLM Gateway. People who leave usually cite one of four things.
1. Seats. On the Plus plan you pay per seat, then pay as you go for usage in LangChain Standard Units (LSUs). The pricing FAQ says a seat is "a distinct user inside your organization" and that LangChain counts every user, invited users included. A team where PMs and support staff only read traces still pays for each of them.
2. Self-hosting is an Enterprise add-on. Self-hosted LangSmith "is an add-on to the Enterprise plan designed for LangChain's largest, most security-conscious customers." If your compliance team says traces can't leave your network and you don't have an Enterprise budget, you're stuck.
3. Retention. LangSmith's pricing FAQ says base traces are kept for 14 days and extended traces for 180 days, and that upgrading a trace to extended costs extra. If you want to look back at last quarter's failures, that adds up.
4. LangChain gravity. Tracing itself isn't locked in. LangSmith accepts OpenTelemetry traces "from any OpenTelemetry-compatible agent," and integrates with OpenAI, Anthropic, CrewAI and Pydantic AI. The pull is in the extras. Engine, for example, can open pull requests only for agents built with Deep Agents, LangChain and LangGraph. If you're moving off LangChain, or never used it, you pay for features you can't fully use.
Quick comparison
| Tool | Fixes which LangSmith pain | Licence | Free self-host? | Cloud billing shape |
|---|---|---|---|---|
| LangSmith (baseline) | — | Proprietary | No, Enterprise add-on | Per seat plus LSU usage |
| Langfuse | Seats, self-hosting | MIT core, ee folders separate | Yes | Usage units, unlimited users on paid tiers |
| Arize Phoenix | Self-hosting, ops weight | Elastic License 2.0 | Yes | Managed version is Arize AX |
| Opik | Self-hosting, licence | Apache-2.0 | Yes, full platform | Comet cloud |
| MLflow | Self-hosting, consolidation | Apache-2.0 | Yes | Managed on Databricks and others |
| OpenLLMetry | Vendor-neutral instrumentation | Apache-2.0 | It's a library, not a backend | Send to any OTel backend |
| Pydantic Logfire | Whole-app OTel view | MIT SDK, closed platform | No, enterprise licence | Usage-based with seats per plan |
| Braintrust | Seats, evals | Proprietary | No for new customers; BYOC instead | Usage-based, unlimited users |
| PromptLayer | Prompt workflow | Proprietary | No, Enterprise only | Plans with user and request limits |
If self-hosting is the problem
These four run on your own infrastructure with no licence key and no sales call.
Langfuse
Langfuse is the most common landing spot for LangSmith leavers, because it covers the same core: tracing, LLM-as-judge and human evals, datasets, a playground, and prompt management with versions and labels. It's built on OpenTelemetry, and its LangChain integration traces LangChain and LangGraph apps through a callback handler, so you can keep your framework and only swap the observability layer.
Self-hosting means running Postgres, ClickHouse, Redis and blob storage, via Docker Compose for testing or Kubernetes for production. That's real ops work, but it's free. Langfuse is also the answer to seat pain: its pricing page lists unlimited users on its paid cloud tiers and bills by ingested units instead.
One ownership note: Langfuse joined ClickHouse in January 2026 and said there are no planned licensing changes. Our Langfuse alternatives roundup covers that, and what to pick if it bothers you.
Arize Phoenix
Phoenix is the lightest self-host on this list. pip install arize-phoenix and phoenix serve give you a working instance, and there's a Docker image if you prefer. It traces with "OpenTelemetry-based instrumentation," runs response and retrieval evals, versions datasets and tracks experiments. The README lists out-of-the-box support for the OpenAI Agents SDK, Claude Agent SDK, LangGraph, Vercel AI SDK, CrewAI, LlamaIndex and DSPy.
The licence is the Elastic License 2.0, which is source-available rather than OSI open source. For internal use that rarely matters. It does matter if you planned to resell it as a hosted service.
Opik
Opik from Comet is Apache-2.0, and its README is explicit that you're "free to self-host the full platform," backend included. It covers tracing, datasets, experiments, LLM-as-judge and online evaluation, and prompt management. It says it's framework-agnostic and offers first-party OpenTelemetry support, so any language with an OTel SDK can send traces. Docker Compose handles local use; Helm handles production.
MLflow
If your team already tracks ML experiments in MLflow, the shortest path is to keep LLM traces there too. MLflow Tracing describes itself as "a fully OpenTelemetry-compatible LLM observability solution," and MLflow's GenAI features add LLM-as-judge evaluation and a prompt registry. It's Apache-2.0, and you can self-host it or use a managed version on Databricks and other clouds.
If you want OpenTelemetry and no vendor in the middle
Some teams leave LangSmith because they don't want any single platform owning their instrumentation. These two emit standard OpenTelemetry, so the backend becomes a config choice.
OpenLLMetry
OpenLLMetry is "a set of extensions built on top of OpenTelemetry" that instruments LLM providers and vector databases. Because the output is plain OTel, it can feed tools you already run, such as Datadog or Honeycomb, or any of the LLM backends in this post that accept OTLP. It's Apache-2.0.
It isn't a dashboard. You still need somewhere to send the spans. Also note the ownership change: Traceloop, the company behind it, announced it's joining ServiceNow in March 2026. The post says OpenLLMetry "will remain open-source," and the repository was still getting commits this week.
Pydantic Logfire
Logfire calls itself "an opinionated wrapper around OpenTelemetry." It watches your whole application, not only LLM calls, so a slow database query and a slow model call show up in the same trace. The SDKs (Python, TypeScript, Rust) are open source and can export to any OTel backend. The platform itself is closed source; self-hosting it requires an enterprise licence.
It's the natural pick for Python teams, especially on Pydantic AI, who want one tool for general and LLM observability.
If evals or prompts were the reason you used LangSmith
Braintrust
Braintrust is built around evaluation: experiments that compare prompts and models, scoring, and an AI assistant that helps you write evals. On billing, Braintrust's pricing lists unlimited users and charges on processed data and scores, so headcount doesn't drive the bill.
Check the hosting story before you pick it for data-residency reasons. Braintrust's self-hosting docs now say self-hosting is a legacy option that, starting November 4, 2026, "isn't available to new customers." New customers who need data in their own cloud are pointed to Bring Your Own Cloud (BYOC) instead.
PromptLayer
PromptLayer pitches itself as "the collaboration layer for AI engineering teams": a prompt registry, evals and observability, built so domain experts can edit and test prompts without touching your codebase. If your LangSmith usage was mostly the Prompt Hub and playground, this is the closest fit.
It has an OpenTelemetry endpoint, so an existing OTel pipeline can send spans without the PromptLayer SDK. Self-hosting is "an Enterprise-only feature," the same limit LangSmith has.
What you give up when you leave
Be honest about this before you migrate. LangSmith's platform extras don't have direct equivalents in most of the tools above:
| LangSmith feature | What replaces it |
|---|---|
| Tracing and evals | Any tool in this post |
| Prompt Hub and playground | Langfuse, Opik, MLflow, PromptLayer |
| LangSmith Deployment (managed agent runtime) | Nothing here. You host the agent yourself |
| Engine (automated issue triage and fix PRs) | Partial overlap at best |
| Fleet, Sandboxes, LLM Gateway | Separate products, if you need them |
The good news: you can keep LangGraph and still leave LangSmith. Langfuse, Phoenix, Opik and MLflow all trace LangChain and LangGraph apps. If you're also rethinking the framework, read LangGraph vs LangChain first. Observability should follow the framework choice, not lead it.
How to choose a LangSmith alternative
| Your situation | Shortlist |
|---|---|
| Traces must stay on your servers, no Enterprise budget | Langfuse, Phoenix, Opik, MLflow |
| Many read-only viewers, seat cost hurts | Langfuse, Braintrust |
| You want to self-host on a laptop or one VM | Phoenix |
| You want standard OTel and freedom to switch backends | OpenLLMetry plus any OTLP backend |
| Python app, want LLM and non-LLM traces together | Logfire |
| Evals drive your release process | Braintrust, Opik |
| Non-engineers own the prompts | PromptLayer, Langfuse |
| You rely on LangSmith Deployment | Stay, or plan to host agents yourself |
Three practical tips:
- Switch instrumentation first. Move to OpenTelemetry spans while you're still on LangSmith, since it accepts them. Then changing the backend is a config change.
- Run two tools side by side for a sprint. Most of these have a free cloud tier or a free self-host. Compare them on your real traces, not a demo.
- Price the ops time. Phoenix is one process. Langfuse is four storage services. The licence is free in both cases; your hours aren't.
For current prices, read each vendor's pricing page yourself. Billing models change often.
Where Taku fits
Everything above is for teams who build LLM apps and need to see inside them. Taku doesn't trace or evaluate anything. Taku is for the people on the other side: an AI-native desktop workspace where you mirror an AI app, skill or workflow someone already got working, run it, remix it, and keep it. Taku is in Beta, and the Mac app is available now.
FAQ
What is the best open source LangSmith alternative?
Langfuse is the most common choice, because it covers tracing, evals and prompt management and you can self-host it for free. Opik is the pick if you want an Apache-2.0 licence on the whole platform. Phoenix is the easiest to run, though its Elastic License 2.0 is source-available rather than OSI open source.
Can I self-host LangSmith?
Only as an add-on to the Enterprise plan. LangChain's docs say you contact sales for a licence key. If that's not an option, Langfuse, Phoenix, Opik and MLflow can all be self-hosted without one.
Does LangSmith only work with LangChain?
No. It accepts OpenTelemetry traces from any compatible app and integrates with OpenAI, Anthropic, CrewAI and others. Some extras are narrower, though. Engine's fix PRs, for example, cover agents built with Deep Agents, LangChain and LangGraph.
Is there a LangSmith alternative without per-seat pricing?
Yes. Langfuse's paid cloud tiers include unlimited users and bill by usage, and Braintrust's pricing also lists unlimited users. Self-hosting an open-source tool removes seat costs entirely, in exchange for running it yourself.
Can I keep LangGraph and switch observability tools?
Yes. Langfuse, Phoenix, Opik and MLflow all trace LangChain and LangGraph apps. What you lose is LangSmith Deployment, so you'd host the agent somewhere else.
Which LangSmith alternatives support OpenTelemetry?
Langfuse, Phoenix, Opik, MLflow, OpenLLMetry, Logfire and PromptLayer all build on or accept OpenTelemetry. Standard spans keep you free to switch again later.