Which AI Agent Framework Do People Install? 155,918,257 LangChain Downloads in 30 Days

155,918,257. That is how many times the PyPI distribution langchain was downloaded from October 10 through November 8, 2026, in the without_mirrors series.

The November 5 star snapshot is the page that already shows up for agent-framework stats. It ranks repositories by stars and latest release. It does not say which distribution pip actually fetched. This page sums the daily download rows for the distributions behind that set, adds the one the snapshot left out, and reads the star field again on November 9.

Which AI agent framework do people install?

langchain, by PyPI file downloads, in this set and this window. langgraph is second. strands-agents is third, ahead of claude-agent-sdk and openai-agents.

The window is 30 calendar days, October 10 through November 8, inclusive. Each total is the sum of the daily without_mirrors counts from https://pypistats.org/api/packages/{name}/overall?mirrors=false. The series for these packages ended on November 8. November 9 was not in the file yet. Stars are stargazers_count from the GitHub repository API. The responses for the repositories that were already in the November 5 set carried HTTP Date headers from 13:38:56Z to 13:39:03Z on November 9, 2026. strands-agents/harness-sdk was read at 13:40:17Z, because the PyPI homepage pointed there and the first list did not include it.

The 19 thirty-day totals sum to 365,081,875. That sum double-counts work that installs more than one of these distributions. It is a checksum on the rows, and it is not a market size.

pydantic-ai 2.66.0 requires pydantic-ai-slim==2.66.0 with no extra marker. The slim distribution had 32,785,681 downloads, 4.99 times the meta package’s 6,575,730. Both rows share the star count of pydantic/pydantic-ai. The directory page is the product. The two download rows are two files on PyPI.

agent-framework 1.25.0 requires agent-framework-core[all]==1.25.0 with no extra marker. Core had 1,469,689 downloads and the meta package had 195,287. Both point at microsoft/agent-framework, which had 14,603 stars, so the star cell is one repository printed twice. The directory listing is that repository.

The order of the top five distributions is the same in the seven-day column as in the thirty-day column: langchain, langgraph, strands-agents, claude-agent-sdk, pydantic-ai-slim.

A cross-check on the same endpoint: for langgraph, the seven daily rows from November 2 through November 8 sum to 13,876,217, and the recent endpoint’s last_week field was 13,876,217. November 8 alone was 1,109,424, and last_day was 1,109,424. The recent endpoint’s last_month field was 44,769,952. The thirty daily rows sum to 55,093,579. The same split showed up for langchain (170,457,766 in last_month, 155,918,257 in the thirty daily rows) and for crewai (2,442,764 against 2,629,769). This page uses the daily rows.

Why does CrewAI have more stars and fewer downloads?

CrewAI’s repository had 15,995 more stars than LangGraph’s. LangGraph’s 30-day download count was 20.95 times CrewAI’s.

60,684 minus 44,689 is 15,995. 55,093,579 divided by 2,629,769 is 20.95. The seven-day cut is the same shape: 13,876,217 against 590,729.

Stars are an account’s bookmark on a repository. Downloads are file requests for one named distribution. CrewAI’s PyPI summary calls it a framework for orchestrating role-playing, autonomous agents, and the project URL is crewAIInc/crewAI. LangGraph’s summary is “Building stateful, multi-actor applications with LLMs,” and the project URL is the libs/langgraph tree inside langchain-ai/langgraph. The CrewAI listing and the LangChain listing describe the products. The framework comparison is the licence and release table. Neither of those pages is an install ranking.

langchain itself had 59.29 times CrewAI’s downloads, 155,918,257 against 2,629,769, and 149,260 stars against 60,684. The November 5 snapshot had this repository at 149,047 stars. Four days later the field was 213 higher. The download window is the new measurement.

Where does strands-agents rank if the star table left it out?

Third on this PyPI list. 37,592,185 downloads. 9,084 stars on the repository the package names.

The November 5 methodology says strands-agents was left out because the GitHub API returned a redirect. On November 9, PyPI’s JSON for strands-agents 1.64.0 gave the summary “A model-driven approach to building AI agents in just a few lines of code,” Apache-2.0 in the license field, homepage https://github.com/strands-agents/harness-sdk, and documentation https://strandsagents.com. The API for that repository returned 9,084 stars, 1,487 forks, the SPDX id Apache-2.0, archived: false, and the description “Build an agent harness and control it end-to-end. Open-source SDK for production AI agents in Python & TypeScript - any model, any cloud.”

37,592,185 divided by 2,629,769 is 14.29. The omitted package had 14.29 times CrewAI’s downloads and 9,084 stars against CrewAI’s 60,684. A harness, on this site, is the loop, tools, memory, permissions, and stop rules around a model. The repository description uses that word for the SDK. This page did not install it.

claude-agent-sdk is the next row, 33,887,233 downloads, from anthropics/claude-agent-sdk-python at 8,502 stars. Version 0.2.179. The PyPI summary is “Python SDK for Claude Code.” The listing is the product page. openai-agents 0.24.0 had 17,426,915 downloads and the Python repository had 30,573 stars. That listing is the OpenAI Agents SDK. google-adk 2.14.0 had 10,099,299 downloads and google/adk-python had 22,158 stars. Its listing is Google ADK.

Does installing LangChain also download LangGraph?

A fresh install of langchain 1.4.9 requests langgraph. The 30-day totals still do not nest, because 155,918,257 is larger than 55,093,579.

The requires line, read at 13:40:16 UTC, is langgraph<1.3.0,>=1.2.12, with no extra marker. A cache that already holds the file, and every older langchain release inside the 30-day total, keeps the two numbers from lining up as parent and child. google-adk’s langgraph lines are different: each one is guarded by extra == "all", extra == "extensions", or extra == "test". A default pip install google-adk does not, on that metadata, require langgraph.

autogen-agentchat 0.7.5 had 401,015 downloads. The repository microsoft/autogen had 61,759 stars on this pass. langchain’s download total was 388.8 times that package’s. The same repository’s README, and the archive flag, are the subject of Is AutoGen deprecated?, which read 61,753 stars at 07:41 UTC the same day. Six stars later, the install count is still the small number. semantic-kernel 1.47.0 lists autogen-agentchat only behind extra == "autogen" and a Python-version marker. That extra is not the 401,015.

What do the npm packages show for the same dates?

The package ai had 140,196,706 downloads. @langchain/langgraph had 15,932,146. @mastra/core had 9,988,333.

npm’s point API was called for the range 2026-10-10:2026-11-08. Each response echoed that start and that end. The download-counts doc describes this endpoint. An npm download and a PyPI download are different registries. This table does not add them to the PyPI column.

The registry describes ai as the AI SDK by Vercel, a single interface for model providers. The GitHub description for vercel/ai, trailing space removed, was “The AI Toolkit for TypeScript. From the creators of Next.js, the AI SDK is a free open-source library for building AI-powered applications and agents.” 140,196,706 downloads and 27,830 stars sit on that repository. The November 5 snapshot included vercel/ai in its 19. The download count covers every app that depends on the package, including apps that call a model and never run an agent loop.

@langchain/langgraph names langchain-ai/langgraphjs, which had 3,411 stars at 13:42:28Z. The Python repository langchain-ai/langgraph had 44,689 stars and the PyPI distribution had 55,093,579 downloads. Those are two repositories. @openai/agents names openai/openai-agents-js (3,996 stars), and the Python SDK is the other repository (30,573 stars, 17,426,915 PyPI downloads). @google/adk names google/adk-js (1,471 stars, version 2.4.0). The Python package google-adk was on 2.14.0. @mastra/core 1.82.0 names mastra-ai/mastra, whose description was “Mastra is the modern TypeScript framework for AI-powered applications and agents,” at 29,428 stars. The npm metadata did not include a description field for that package. The star pass’s HTTP Date for Mastra was 13:39:01Z. The three smaller JavaScript repositories were read at 13:42:28Z and 13:42:29Z.

Why is Letta absent from the table?

The PyPI package letta and the repository in the November 5 star table are different projects, so this page does not pair them.

letta 0.34.12 had 9,886 downloads in the same 30 days. Its summary was “Letta Code: stateful agents in your terminal,” and its source URL was https://github.com/letta-ai/letta-code. The star snapshot’s row is letta-ai/letta, which returned 25,390 stars, pushed_at 2026-10-18T22:28:04Z, and the SPDX id Apache-2.0. Pairing 9,886 downloads with 25,390 stars would answer a question about the wrong repository.

What is a download count for?

A download count is the right tool for “what did pip and npm fetch,” and a weak tool for “what should I install.”

CI, a dependency, and a person trying a tutorial all increment the same field. langchain 1.4.9 pulling langgraph, and pydantic-ai pulling pydantic-ai-slim, are the cases checked above. Other dependencies inside these graphs were not fully expanded. Quality, licence, and whether the README still promises fixes are separate questions. AutoGen’s install name in its own README is autogen-agentchat, and that distribution’s 401,015 downloads sit next to a repository whose maintenance statement is already written up.

The agent picker is the directory’s filter for a job. How do AI agents work? is the loop those libraries wrap. A star census of agent products, rather than of these libraries, is How many GitHub stars does OpenClaw have?. How many popular repositories write instructions for the coding agents these SDKs drive is How many repos use AGENTS.md?.

FAQ

Which AI agent framework do people install?

langchain, at 155,918,257 PyPI downloads from October 10 through November 8, 2026, without mirrors. langgraph had 55,093,579. On npm, over the same dates, the package ai had 140,196,706. The two registries are not one ranking.

Does CrewAI have more stars than installs would suggest?

It has more stars than LangGraph and fewer downloads. 60,684 stars against 44,689. 2,629,769 downloads against 55,093,579. The download ratio is 20.95.

Why was strands-agents left off the star snapshot?

The snapshot’s note says the GitHub API returned a redirect. The package on PyPI now points at strands-agents/harness-sdk, which had 9,084 stars on November 9 and 37,592,185 downloads in the 30-day window.

Does the LangChain number include LangGraph?

Installing langchain 1.4.9 requests langgraph. The published totals do not nest: 155,918,257 is larger than 55,093,579. Older releases and caches sit inside the larger number.

Is a download a user?

A download is a file request in one registry’s public stats. It includes CI and dependency installs. It is the count to use when the question is what was fetched.