langchain-ai/langchain
The agent engineering platform.
How langchain is put together
This is the LangChain Python monorepo, a set of composable libraries for building LLM applications. It is structured as a multi-package workspace under libs/: 'langchain-core' provides foundational abstractions (messages, runnables, caches, retrievers, API lifecycle decorators), 'langchain' (libs/langchain_v1) is the modern v1 package built on langgraph, 'langchain-classic' (libs/langchain) retains legacy chains/agents/retrievers, plus text-splitters, standard-tests, model-profiles CLI, and partner integration packages under libs/partners/. All packages are pure Python (>=3.10) built with hatchling and managed with uv; there are no servers or datastores in this repo itself — it ships libraries consumed by downstream applications. Data flow is message/prompt objects through runnables/chains to provider SDKs in partner packages.
Languages
Frameworks
Infrastructure
Major components
langchain-core (libs/core/langchain_core)
Foundational abstractions: messages, prompt values, runnables, caches, retrievers, stores, rate limiters, exceptions, and API lifecycle decorators (deprecation/beta).
langchain v1 (libs/langchain_v1/langchain)
The current 'langchain' package providing agent abstractions on top of langchain-core and langgraph.
langchain-classic (libs/langchain/langchain_classic)
Legacy package retaining classic chains, memory, agents, SQL database utilities, text splitters, and backwards-compatibility import shims.
langchain-text-splitters (libs/text-splitters/langchain_text_splitters)
Text splitting implementations (character, HTML, JSON, JSX) used to chunk documents for retrieval.
langchain-tests / standard-tests (libs/standard-tests/langchain_tests)
Shared pytest test suites and a pytest11 plugin (_langsmith_plugin) that integration packages use to validate their implementations.
partner integrations (libs/partners/*)
Provider-specific packages (openai, anthropic, groq, mistralai, ollama, qdrant, chroma, huggingface, etc.) each with its own pyproject/uv.lock/Makefile.
model-profiles CLI (libs/model-profiles/langchain_model_profiles)
Console tool ('langchain-profiles') for updating model profile data in LangChain integration packages.
_api lifecycle module (langchain_core._api)
Internal helpers for @deprecated/@beta decorators, warning suppression context managers, and caller-internality detection.
A quieter week centered on a new LangChain release, a fix for human-in-the-loop rejection reasons, and routine dependency updates.
5 weeksWeek of 2026-08-24
Week of 2026-08-24latest
A quieter week centered on a new LangChain release, a fix for human-in-the-loop rejection reasons, and routine dependency updates.
Week of 2026-08-17
A very busy week with major releases of core, OpenAI, and Anthropic packages, plus dozens of fixes around tool calls, streaming, and error handling.
Week of 2026-08-10
OpenAI SDK 3.0 support landed alongside a wave of fixes for streaming, caching, and chunk merging across providers.
Week of 2026-08-03
Fifty-plus changes focused on hardening: better error messages, safer handling of malformed responses, and fixes across middleware and provider integrations.
Week of 2026-07-27
Security and correctness took center stage, with credential redaction, preservation of caller settings, and a series of core tool-handling fixes.
Over the past month the LangChain team shipped a steady stream of releases across its core library and AI provider integrations (OpenAI, Anthropic, Fireworks, Perplexity, OpenRouter), with a heavy focus on fixing how tool calls, streaming responses, and usage metadata are handled. Security and reliability work was prominent: credentials are now redacted from logs and errors, malformed model responses are handled gracefully, and dependencies were regularly bumped. New capabilities included document reranking for Fireworks, support for the OpenAI 3.0 SDK, custom token counting in context editing, and richer gateway metadata surfaced to users.
Week by week
2026-08-24A quieter week centered on a new LangChain release, a fix for human-in-the-loop rejection reasons, and routine dependency updates.latest6 changes
Chore
LangChain 1.3.17 released
A new patch release of the main LangChain package rolled up recent fixes.
Fix
Custom human-in-the-loop rejection reasons fixed
Fixed how custom reasons are framed when a human reviewer rejects an agent's proposed action.
Feature
Anthropic gateway metadata surfaced
Responses routed through a gateway now include extra response metadata so users can see routing details.
Chore
Model profile data refreshed
The catalog of known AI model characteristics was updated with the latest information.
Chore
Python base image upgraded
Docker builds for LangChain moved from Python 3.11 to Python 3.14 as their foundation.
Docs
Corridor agent instructions added
Documentation was added describing how agents should operate in the Corridor environment.
2026-08-17A very busy week with major releases of core, OpenAI, and Anthropic packages, plus dozens of fixes around tool calls, streaming, and error handling.6 changes
Feature
Standard model exception types added
Core now defines a standard set of error types that all model integrations raise, making failures easier to catch and handle.
Feature
Fireworks document reranking added
Fireworks integration gained the ability to rerank documents by relevance.
Fix
Invalid tool calls filtered from content
Both Fireworks and Anthropic integrations now strip out broken or incomplete tool-call blocks before returning results.
Fix
Strict tool schemas require nested properties
Tool definitions sent to models now correctly mark all nested fields as required under strict mode.
Fix
OpenAI reasoning item boundaries preserved
Streaming responses no longer merge separate reasoning steps from OpenAI models into one blob.
Refactor
Transformers import made lazy
A heavyweight machine-learning library is now only loaded when actually needed, speeding up startup.
2026-08-10OpenAI SDK 3.0 support landed alongside a wave of fixes for streaming, caching, and chunk merging across providers.6 changes
Feature
OpenAI 3.0 SDK supported
The OpenAI integration was upgraded to work with version 3.0 of OpenAI's official client library.
Feature
Trace policy exposed on agent middleware
Developers can now control which middleware calls get recorded in traces when debugging agents.
Fix
Encrypted reasoning preserved in streams
OpenAI's encrypted reasoning data is no longer lost when responses arrive as a stream.
Fix
Pydantic aliases respected in tool inputs
Tool inputs using alternate field names (aliases) are now validated correctly.
Fix
Reasoning tokens reported in Anthropic usage
Token counts from Anthropic models now include the tokens used for internal reasoning.
Fix
Cost metadata kept in OpenRouter usage chunks
Pricing information returned by OpenRouter is no longer dropped from streamed usage reports.
2026-08-03Fifty-plus changes focused on hardening: better error messages, safer handling of malformed responses, and fixes across middleware and provider integrations.6 changes
Fix
Summarization middleware preserves history on failure
If generating a conversation summary fails, the original chat history is kept intact instead of being lost.
Fix
Malformed structured-output responses handled
Agents no longer crash when a model returns badly formatted structured output.
Fix
ContextWindowExceededError handled in OpenAI
Hitting a model's context limit now produces a clear, catchable error rather than an obscure failure.
Feature
state_schema parameter added to wrap_tool_call
Custom tool wrappers can now declare what parts of agent state they need access to.
Fix
Pydantic 2.14 compatibility restored
Core was updated to keep working with the newest version of the Pydantic data-validation library.
Fix
Clearer error when Anthropic credentials missing
Forgetting to configure API keys for Claude now yields an obvious, helpful error message.
2026-07-27Security and correctness took center stage, with credential redaction, preservation of caller settings, and a series of core tool-handling fixes.6 changes
Fix
Streaming callback options redacted
Sensitive configuration values passed during streaming are now hidden from logs and callbacks.
Fix
MCP authorization headers redacted
Authentication tokens sent to MCP servers are no longer exposed in OpenAI request logs.
Fix
Caller tool_choice preserved in Anthropic
Settings controlling how Claude picks tools are no longer overwritten by the integration.
Fix
Gateway falls back to LangSmith API key
If no dedicated gateway key is set, the system can reuse an existing LangSmith key instead of failing.
Feature
user_profile_id convenience attribute added
Anthropic users get a simpler way to attach a user profile identifier to requests.
Chore
LICENSE files added to publishable packages
All published packages now ship with their license files included, as required by package registries.
Dependencies and code review
Dependency advisories
Security Watch
No known advisories across 0 scanned dependencies.
No known advisories in the scanned dependencies.
Code review
No concrete code-level issues confirmed.
Get this report every week for your repos.
GitZoid learns each repo, reports what changed, and flags what needs attention. One flat price for the whole team.
$19 a month, flat · First 10 outputs free · No card required