LangSmith

0.7.7 · active · verified Sat Feb 28

Official Python SDK for LangSmith — LangChain's observability, tracing, and evaluation platform. Instruments LLM calls, chains, agents, and arbitrary functions with the @traceable decorator or wrap_openai(). Traces are sent asynchronously to LangSmith cloud (or self-hosted). Also provides dataset management and evaluation (evaluate()) APIs. Usable standalone without LangChain. Releases multiple times per week. SECURITY NOTE: CVE-2026-25528 — SSRF via baggage header injection, fixed in 0.6.3. Any version 0.4.10–0.6.2 is vulnerable.

Warnings

breaking CVE-2026-25528: SSRF via baggage header injection. The distributed tracing feature is vulnerable — attackers can inject arbitrary api_url values via the baggage HTTP header, causing the SDK to exfiltrate trace data to attacker-controlled endpoints. Affects versions 0.4.10–0.6.2.
Fix: Upgrade immediately: pip install 'langsmith>=0.6.3'. Do not run affected versions in environments that process untrusted HTTP requests.
breaking LANGCHAIN_TRACING_V2 and LANGCHAIN_API_KEY are the legacy env var names. Current docs use LANGSMITH_TRACING and LANGSMITH_API_KEY. Both still work, but mixing old and new names in the same environment causes confusing precedence issues.
Fix: Standardize on LANGSMITH_TRACING and LANGSMITH_API_KEY in all new code and CI configs. Remove LANGCHAIN_TRACING_V2 from environments.
breaking langsmith 0.5+ has a strict version conflict with older langchain versions. langchain 0.3.x requires langsmith<0.2.0. Installing langsmith>=0.5 in a langchain 0.3 environment causes a pip resolver conflict that breaks the entire environment.
Fix: Upgrade langchain to the latest version before upgrading langsmith. Or pin both: langchain==0.3.x requires langsmith<0.2.
gotcha LANGSMITH_TRACING must be set BEFORE LangChain is imported. LangChain reads the env var at import time. Setting os.environ['LANGSMITH_TRACING'] = 'true' after 'from langchain...' has already run produces zero traces with no error.
Fix: Set all LANGSMITH_* env vars at the very top of your entry point, before any LangChain/LangSmith imports. Use .env files loaded via python-dotenv before other imports.
gotcha LANGSMITH_TRACING='true' must be a string, not a boolean. In some YAML/env file parsers, setting LANGSMITH_TRACING: true (no quotes) passes the Python boolean True, which the SDK does not recognize. Traces silently drop.
Fix: Always quote the value: LANGSMITH_TRACING='true' (shell) or LANGSMITH_TRACING: 'true' (YAML).
gotcha Traces are sent asynchronously in a background thread. In short-lived scripts, the process may exit before all traces are flushed, resulting in missing or partial traces in the UI.
Fix: Call langsmith.utils.wait_for_all_tracers() at the end of short-lived scripts to ensure all traces are flushed before exit.

Install

pip install langsmith Core SDK
pip install 'langsmith[otel]' With OpenTelemetry export support
pip install 'langsmith[pytest]' With pytest integration for evals

Imports

traceable
```
from langsmith import traceable
```
Primary decorator for instrumenting any function. Works standalone — does not require LangChain.
wrap_openai
```
from langsmith.wrappers import wrap_openai
```
Wraps an OpenAI client to auto-trace all calls. Equivalent wrappers exist for Anthropic (wrap_anthropic).
Client
```
from langsmith import Client
```
Low-level client for dataset management, run querying, and evaluation APIs.

Quickstart

Env vars must be set BEFORE importing LangChain/LangSmith — they are read at import time. Setting them after import has no effect. LANGSMITH_TRACING must be the string 'true', not a boolean.

import os
os.environ['LANGSMITH_TRACING'] = 'true'
os.environ['LANGSMITH_API_KEY'] = 'ls_...'
os.environ['LANGSMITH_PROJECT'] = 'my-project'

# Option 1: Wrap OpenAI client (auto-traces all calls)
import openai
from langsmith.wrappers import wrap_openai

client = wrap_openai(openai.OpenAI())
response = client.chat.completions.create(
    model='gpt-4o',
    messages=[{'role': 'user', 'content': 'Hello!'}]
)

# Option 2: @traceable decorator for arbitrary functions
from langsmith import traceable

@traceable
def my_pipeline(query: str) -> str:
    # any code here is traced as a span
    response = client.chat.completions.create(
        model='gpt-4o',
        messages=[{'role': 'user', 'content': query}]
    )
    return response.choices[0].message.content

result = my_pipeline('What is 2+2?')

# Option 3: LangChain auto-tracing (no decorator needed)
# Just set env vars — all LangChain/LangGraph calls are traced automatically
from langchain_openai import ChatOpenAI
llm = ChatOpenAI(model='gpt-4o')
llm.invoke('Hello!')  # automatically traced

view raw JSON →