The open-source LangSmith alternative

Q: Is there an open-source alternative to LangSmith?

Yes. Langprobe is an open-source (Apache-2.0), self-hosted alternative to LangSmith. It provides tracing, evals, and monitoring like LangSmith, and adds replay debugging and an agent-native MCP surface. Langfuse (MIT) is another open-source option; unlike both LangSmith and Langfuse, Langprobe lets you edit and re-run a broken run against a real model and diff it span by span.

Q: Can I self-host LangSmith for free?

LangSmith is a proprietary SaaS from LangChain; self-hosting it requires an Enterprise plan. Langprobe is Apache-2.0 and self-hosts for free with one docker compose up — Postgres, ClickHouse, Redis, and S3/MinIO — in your own VPC.

Q: Do I need the LangChain framework to use Langprobe?

No. Langprobe ingests plain OTLP/HTTP via OpenTelemetry and OpenInference instrumentors, so it works with CrewAI, DSPy, Pydantic AI, OpenAI Agents, LlamaIndex, LangChain, or bare provider SDKs. LangSmith is most tightly coupled to the LangChain ecosystem.

Langprobe gives you LangSmith-style tracing, evals, and monitoring — as Apache-2.0 open source you self-host in your own VPC — plus two things LangSmith doesn't have: replay debugging and an agent-native MCP surface.

Last updated: July 4, 2026

TL;DR

Langprobe is open-source (Apache-2.0) and free to self-host; LangSmith is a proprietary SaaS whose self-hosting requires an Enterprise plan.
Both do tracing, datasets, evals, and monitoring. Only Langprobe adds replay: edit a prompt/model/tool config on a failed run, re-run it against a real model, and diff span by span with a determinism verdict.
Langprobe exposes a token-budgeted REST + MCP surface so a coding agent can debug an agent; it ingests plain OTLP/HTTP with no proprietary SDK.
Migration is one env var — ingestion is LangSmith-compatible and history backfills automatically.

Langprobe vs LangSmith at a glance

Dimension	Langprobe	LangSmith
License	Apache-2.0 open source	Proprietary / closed source
Free self-hosting	Yes — docker compose up	Enterprise plan only
Data stays in your VPC	Yes, by default	Cloud by default
Tracing & spans	Yes	Yes
Datasets & evals	Yes	Yes
Replay (edit & re-run a broken run)	Yes — span-level diff + determinism verdict	No
Agent-native surface (MCP)	Yes — token-budgeted REST + MCP	REST API; no first-class MCP debug loop
Instrumentation	Open — OTLP/HTTP + OpenInference	Best with LangChain SDK
Framework coupling	Framework-agnostic	LangChain ecosystem

Facts about LangSmith reflect its public positioning as a proprietary SaaS by LangChain with enterprise-gated self-hosting (see Langfuse's comparison and DataCamp's Langfuse vs LangSmith). Verify current terms on LangSmith's site before you commit.

What Langprobe adds that LangSmith doesn't

1. Replay — the 2am debugger

LangSmith shows you what happened. Langprobe lets you change what happened: open a broken run, edit a prompt, model, or tool config, and re-run it against a real model. You get a side-by-side span diff — outputs, latency, tokens, cost — and a determinism verdict that tells you whether a fix is real or just a lucky sample. That's the difference between "I think this is fixed" and "this is fixed 9 times out of 10."

2. An agent can debug an agent

Every read in Langprobe is a token-budgeted, LLM-legible projection over REST and MCP. A 48k-token trace collapses to a ~2k-token salient slice, so a coding agent (Claude, Cursor, an OpenAI Agent) can find → read → replay → diff hands-free. The same API serves humans, CI, and agents.

3. Open by default

Apache-2.0, self-hosted, no proprietary SDK. Your traces — which contain your prompts, user data, and tool I/O — never leave your infrastructure unless you choose.

Choose the right tool

Choose Langprobe if…

You want open source you can read, self-host, and fork (Apache-2.0).
Your data must stay in your own VPC for compliance.
You debug agents and want replay + determinism, not just dashboards.
You want agents and CI to drive the debug loop over MCP/REST.
You use CrewAI, DSPy, Pydantic AI, OpenAI Agents, or LlamaIndex — not only LangChain.

Choose LangSmith if…

You're all-in on the LangChain framework and want the first-party tool.
You prefer a fully managed SaaS and don't need to self-host.
You want the largest incumbent ecosystem and vendor support today.
Replay and an agent-native MCP debug loop aren't priorities.

Migrating from LangSmith takes one env var

Langprobe's ingestion is LangSmith-compatible. Point your existing SDK at your Langprobe host and a migration service backfills historical runs — no application code changes.

- LANGCHAIN_ENDPOINT=https://api.smith.langchain.com
+ LANGCHAIN_ENDPOINT=https://probe.your-company.dev

Read the full LangSmith migration guide →

Frequently asked questions

Is there an open-source alternative to LangSmith?

Yes — Langprobe is open-source (Apache-2.0) and self-hosted. It offers tracing, evals, and monitoring like LangSmith, and adds replay debugging and an agent-native MCP surface. Langfuse (MIT) is another open-source option; unlike both, Langprobe can edit and re-run a broken run against a real model and diff it span by span.

How do I migrate from LangSmith to Langprobe?

Ingestion is LangSmith-compatible: change the endpoint environment variable so your existing SDK points at your Langprobe host, and a migration service backfills historical runs. No app code changes required.

Can I self-host LangSmith for free?

LangSmith self-hosting requires an Enterprise plan. Langprobe self-hosts for free under Apache-2.0 with one docker compose up (Postgres, ClickHouse, Redis, S3/MinIO) in your own VPC.

Do I need the LangChain framework?

No. Langprobe ingests plain OTLP/HTTP via OpenTelemetry/OpenInference, so it works with CrewAI, DSPy, Pydantic AI, OpenAI Agents, LlamaIndex, LangChain, or bare provider SDKs.

Debug your next broken run properly

Open source, in your VPC, Apache-2.0. Free to start.

Start self-hosting Star on GitHub