The LangChain ecosystem is split into focused, independently-versioned packages. Understanding which packages you need — and their version constraints — prevents incompatibilities and keeps upgrades predictable.
Key principles:
- LangChain 1.0 is the current LTS release. Always start new projects on 1.0+. LangChain 0.3 is legacy maintenance-only — do not use it for new work.
- langchain-core is the shared foundation: always install it explicitly alongside any other package.
- langchain-community (Python only) does NOT follow semantic versioning; pin it conservatively.
- LangGraph vs Deep Agents: choose one orchestration approach based on your use case — they are alternatives, not a required stack (see Framework Choice below).
- Provider integrations (model, vector store, tools) are installed separately so you only pull in what you use.
Environment Requirements
| Requirement | Python | TypeScript / Node |
|---|
| Runtime minimum | Python 3.10+ | Node.js 20+ |
| LangChain | 1.0+ (LTS) | 1.0+ (LTS) |
| LangSmith SDK | >= 0.3.0 | >= 0.3.0 |
Framework Choice
Pick **one** agent orchestration layer. You do not need both.
| Framework | When to use | Core extra package |
|---|
| LangGraph | Need fine-grained graph control, custom workflows, loops, or branching | langgraph / @langchain/langgraph |
| Deep Agents | Want batteries-included planning, memory, file context, and skills out of the box | deepagents (depends on LangGraph; installs it as a transitive dep) |
Both sit on top of langchain + langchain-core + langsmith.
Core Packages
Python — always required
| Package | Role | Min version |
|---|
langchain | Agents, chains, retrieval | 1.0 |
langchain-core | Base types & interfaces (peer dep) | 1.0 |
langsmith | Tracing, evaluation, datasets | 0.3.0 |
Python — orchestration (pick one)
| Package | Use when | Min version |
|---|
langgraph | Building custom graphs directly | 1.0 |
deepagents | Using the Deep Agents framework | latest |
Python — model providers (pick the one(s) you use)
| Package | Provider |
|---|
langchain-openai | OpenAI (GPT-4o, o3, …) |
langchain-anthropic | Anthropic (Claude) |
langchain-google-genai | Google (Gemini) |
langchain-mistralai | Mistral |
langchain-groq | Groq (fast inference) |
langchain-cohere | Cohere |
langchain-fireworks | Fireworks AI |
langchain-together | Together AI |
langchain-huggingface | Hugging Face Hub |
langchain-ollama | Ollama (local models) |
langchain-aws | AWS Bedrock |
langchain-azure-ai | Azure AI Foundry |
Python — common tool & retrieval packages
These packages have tighter compatibility requirements — use the latest available version unless you have a specific reason not to.
| Package | Adds | Notes |
|---|
langchain-tavily | Tavily web search (TavilySearch) | Dedicated integration package; prefer latest |
langchain-text-splitters | Text chunking utilities | Semver, keep current |
langchain-community | 1000+ integrations (fallback) | NOT semver — pin to minor series |
faiss-cpu | FAISS vector store (local) | Via langchain-community; use latest |
langchain-chroma | Chroma vector store | Dedicated integration package; prefer latest |
langchain-pinecone | Pinecone vector store | Dedicated integration package; prefer latest |
langchain-qdrant | Qdrant vector store | Dedicated integration package; prefer latest |
langchain-weaviate | Weaviate vector store | Dedicated integration package; prefer latest |
langsmith[pytest] | pytest plugin for LangSmith | Requires langsmith >= 0.3.4 |
langchain-community stability note: This package is NOT on semantic versioning. Minor releases can contain breaking changes. Prefer dedicated integration packages (e.g. langchain-chroma, langchain-tavily) when they exist — they are independently versioned and more stable.
TypeScript — always required
| Package | Role | Min version |
|---|
@langchain/core | Base types & interfaces (peer dep) | 1.0 |
langchain | Agents, chains, retrieval | 1.0 |
langsmith | Tracing, evaluation, datasets | 0.3.0 |
TypeScript — orchestration (pick one)
| Package | Use when | Min version |
|---|
@langchain/langgraph | Building custom graphs directly | 1.0 |
deepagents | Using the Deep Agents framework | latest |
TypeScript — model providers (pick the one(s) you use)
| Package | Provider |
|---|
@langchain/openai | OpenAI (GPT-4o, o3, …) |
@langchain/anthropic | Anthropic (Claude) |
@langchain/google-genai | Google (Gemini) |
@langchain/mistralai | Mistral |
@langchain/groq | Groq (fast inference) |
@langchain/cohere | Cohere |
@langchain/aws | AWS Bedrock |
@langchain/azure-openai | Azure OpenAI |
@langchain/ollama | Ollama (local models) |
TypeScript — common tool & retrieval packages
| Package | Adds | Notes |
|---|
@langchain/tavily | Tavily web search (TavilySearch) | Dedicated integration package; prefer latest |
@langchain/community | Broad set of community integrations | Use sparingly; prefer dedicated packages |
@langchain/pinecone | Pinecone vector store | Dedicated integration package; prefer latest |
@langchain/qdrant | Qdrant vector store | Dedicated integration package; prefer latest |
@langchain/weaviate | Weaviate vector store | Dedicated integration package; prefer latest |
@langchain/core must be installed explicitly in yarn workspaces and monorepos — it is a peer dependency and will not always be hoisted automatically.
Minimal Project Templates
Minimal dependency set for a LangGraph project (provider-agnostic).
# requirements.txt
langchain>=1.0,<2.0
langchain-core>=1.0,<2.0
langgraph>=1.0,<2.0
langsmith>=0.3.0
# Add your model provider, e.g.:
# langchain-openai
# langchain-anthropic
# langchain-google-genai
Minimal package.json dependencies for a LangGraph project (provider-agnostic).
{
"dependencies": {
"@langchain/core": "^1.0.0",
"langchain": "^1.0.0",
"@langchain/langgraph": "^1.0.0",
"langsmith": "^0.3.0"
}
}
Minimal dependency set for a Deep Agents project (provider-agnostic).
# requirements.txt
deepagents # bundles langgraph internally
langchain>=1.0,<2.0
langchain-core>=1.0,<2.0
langsmith>=0.3.0
# Add your model provider, e.g.:
# langchain-anthropic
# langchain-openai
Minimal package.json dependencies for a Deep Agents project (provider-agnostic).
{
"dependencies": {
"deepagents": "latest",
"@langchain/core": "^1.0.0",
"langchain": "^1.0.0",
"langsmith": "^0.3.0"
}
}
Adding Tavily search and a vector store to a LangGraph project.
# requirements.txt
langchain>=1.0,<2.0
langchain-core>=1.0,<2.0
langgraph>=1.0,<2.0
langsmith>=0.3.0
# Web search
langchain-tavily # use latest; partner package, semver
# Vector store — pick one:
langchain-chroma # use latest; partner package, semver
# langchain-pinecone # use latest; partner package, semver
# langchain-qdrant # use latest; partner package, semver
# Text processing
langchain-text-splitters # use latest; semver
# Your model provider:
# langchain-openai / langchain-anthropic / etc.
Adding Tavily search and a vector store to a LangGraph project.
{
"dependencies": {
"@langchain/core": "^1.0.0",
"langchain": "^1.0.0",
"@langchain/langgraph": "^1.0.0",
"langsmith": "^0.3.0",
"@langchain/tavily": "latest",
"@langchain/pinecone": "latest"
}
}
Versioning Policy & Upgrade Strategy
| Package group | Versioning | Safe upgrade strategy |
|---|
langchain, langchain-core | Strict semver (1.0 LTS) | Allow minor: >=1.0,<2.0 |
langgraph / @langchain/langgraph | Strict semver (v1 LTS) | Allow minor: >=1.0,<2.0 |
langsmith | Strict semver | Allow minor: >=0.3.0 |
Dedicated integration packages (e.g. langchain-tavily, langchain-chroma) | Independently versioned | Allow minor updates; use latest |
langchain-community | NOT semver | Pin exact minor: >=0.4.0,<0.5.0 |
deepagents | Follow project releases | Pin to tested version in production |
Breaking changes only happen in major versions (1.x → 2.x) for all semver-compliant packages. Deprecated features remain functional across the entire 1.x series with warnings.
Prefer dedicated integration packages over langchain-community. When a dedicated package exists (e.g. langchain-chroma instead of langchain-community's Chroma integration), use it — dedicated packages are independently versioned and better tested.
Community tool packages (Tavily, vector stores, etc.) should be kept at latest unless your project requires a locked environment. These packages frequently release compatibility fixes alongside LangChain/LangGraph updates.
Environment Variables
All keys are read from the environment at runtime. Set only the keys for services you actually use.
# LangSmith (always recommended for observability)
LANGSMITH_API_KEY=<your-key>
LANGSMITH_PROJECT=<project-name> # optional, defaults to "default"
# Model provider — set the one(s) you use
OPENAI_API_KEY=<your-key>
ANTHROPIC_API_KEY=<your-key>
GOOGLE_API_KEY=<your-key>
MISTRAL_API_KEY=<your-key>
GROQ_API_KEY=<your-key>
COHERE_API_KEY=<your-key>
FIREWORKS_API_KEY=<your-key>
TOGETHER_API_KEY=<your-key>
HUGGINGFACEHUB_API_TOKEN=<your-key>
# Common tool/retrieval services
TAVILY_API_KEY=<your-key> # for Tavily search
PINECONE_API_KEY=<your-key> # for Pinecone
Common Mistakes
Never start a new project on LangChain 0.3. It is maintenance-only until December 2026.
# WRONG: legacy, no new features, security patches only
langchain>=0.3,<0.4
# CORRECT: LangChain 1.0 LTS
langchain>=1.0,<2.0
`langchain-community` can break on minor version bumps — it does not follow semver.
# WRONG: allows minor-version updates that may be breaking
langchain-community>=0.4
# CORRECT: pin to exact minor series
langchain-community>=0.4.0,<0.5.0
Also consider switching to the equivalent dedicated integration package if one exists (e.g. langchain-chroma instead of the community Chroma integration).
Community tool packages like `langchain-tavily` and vector store integrations release compatibility fixes alongside LangChain updates. Using an old pinned version can cause import errors or broken tool schemas.
# RISKY: old pin may be incompatible with LangChain 1.0
langchain-tavily==0.0.1
# BETTER: allow latest within the current major
langchain-tavily>=0.1
Many tools that used to live in `langchain-community` now have dedicated packages with updated import paths. Always prefer the dedicated package import.
# WRONG — deprecated community import path
from langchain_community.tools.tavily_search import TavilySearchResults
from langchain_community.tools import WikipediaQueryRun
from langchain_community.vectorstores import Chroma
from langchain_community.vectorstores import Pinecone
# CORRECT — use dedicated package imports
from langchain_tavily import TavilySearch # pip: langchain-tavily (TavilySearchResults is deprecated)
from langchain_community.tools import WikipediaQueryRun # no dedicated pkg yet
from langchain_chroma import Chroma # pip: langchain-chroma
from langchain_pinecone import PineconeVectorStore # pip: langchain-pinecone
To find the current canonical import for any integration, search the integrations directory:
https://python.langchain.com/docs/integrations/tools/
Each entry shows the correct package and import path. If a dedicated package exists, use it — the community path may still work but is considered legacy.
`@langchain/core` is a peer dependency — it must be in your package.json, especially in monorepos.
// WRONG: missing @langchain/core (breaks in yarn workspaces / strict hoisting)
{
"dependencies": {
"@langchain/langgraph": "^1.0.0"
}
}
// CORRECT: always list @langchain/core explicitly
{
"dependencies": {
"@langchain/core": "^1.0.0",
"@langchain/langgraph": "^1.0.0"
}
}
Python 3.9 and below are not supported by LangChain 1.0.
# Verify before installing
import sys
assert sys.version_info >= (3, 10), "Python 3.10+ required for LangChain 1.0"
Node.js below 20 is not officially supported.
# Verify before installing
node --version # must be v20.x or higher
