AI agent framework downloads vs GitHub stars, November 2026: AutoGen is 6.9 per star

6.9. That is autogen-agentchat's PyPI downloads from October 15 through November 13, 2026, divided by microsoft/autogen's star count on the morning of November 14. 425,456 divided by 61,809 is 6.883, printed here as 6.9.

4,249.1 is the other end of the same division. claude-agent-sdk had 36,291,587 downloads. anthropics/claude-agent-sdk-python had 8,541 stars. The pypistats responses carried HTTP Date Sat, 14 Nov 2026 05:17:29 GMT for langchain and 05:17:30 GMT for the rest. The GitHub responses carried 05:17:31 GMT through 05:17:34 GMT.

The downloads tracker is the live file-request table. The stars tracker is the live bookmark table. The November 5 star report ranks repositories and releases. Best AI agent frameworks is the licence table. None of those pages divides the 30-day download total by stars. The November 9 install census prints both columns for an earlier window and stops at the comparison between two packages. This note is the ratio.

Which AI agent framework has the most downloads per GitHub star?

claude-agent-sdk, at 4,249.1. strands-agents is second at 4,156.6. Both sit under 10,000 stars and both cleared 36 million downloads, so the ratio is not a tiny denominator dressed up as a market.

The window is 30 calendar days, October 15 through November 13, inclusive. Each Python total is the sum of daily without_mirrors rows. Stars are stargazers_count from the same morning. Per star is downloads divided by stars, to one decimal. npm is the next section. It is not in this table.

36,291,587 divided by 8,541 is 4,249.1. 38,256,992 divided by 9,204 is 4,156.6. 58,774,245 divided by 44,934 is 1,308.0. 156,867,095 divided by 149,473 is 1,049.5. 425,456 divided by 61,809 is 6.9.

Among repositories with at least 10,000 stars, LangGraph is the high end at 1,308.0 and AutoGen is the low end at 6.9. LangChain has more stars and more downloads than LangGraph, and a lower ratio, 1,049.5, because 149,473 is a large denominator under a large numerator. Raw downloads still put langchain first. Per star does not. Use the column that matches the decision.

PyPI's homepage for strands-agents 1.65.0, read at 05:21:13 GMT, is github.com/strands-agents/harness-sdk. claude-agent-sdk 0.2.182 points at anthropics/claude-agent-sdk-python. Those are the repositories in the star column. The November 14 job picker assigns a job to eleven frameworks and leaves both of these SDKs out of that set. They stay in this table because the download file is where they show up.

Which starred framework has the fewest installs per star?

AutoGen. 61,809 stars, 425,456 downloads, 6.9 per star.

The November 5 report already had microsoft/autogen at 61,706 stars, retrieved at 2026-11-05T04:41:05Z. This morning's count is 103 stars higher. The low ratio is not a repository that got famous this week and has not had time to be installed. smolagents is the next quiet star row at 23.4, on 30,361 stars and 709,647 downloads. LlamaIndex has 52,893 stars and 36.6 downloads per star.

CrewAI is the star-versus-install comparison people already half know, and the ratio makes it sharp. crewAIInc/crewAI had 60,872 stars. langchain-ai/langgraph had 44,934. The gap is 15,938 stars. Downloads run the other way: 2,629,192 against 58,774,245. 58,774,245 divided by 2,629,192 is 22.35. Per star, LangGraph is at 1,308.0 and CrewAI is at 43.2. A star sort picks the crew. A file-request sort picks the graph. The job each one is for is the November 14 picker, not this ratio.

Agno sits with the starred-and-lighter group: 43,183 stars, 2,242,183 downloads, 51.9 per star. agno 3.2.0's Repository field is github.com/agno-agi/agno. AG2 is a different case. ag2ai/ag2 has only 5,063 stars, and ag2 had 215,889 downloads, 42.6 per star. That is close to CrewAI's ratio on a repository one twelfth the size. Do not read 42.6 as "AG2 is as widely fetched as CrewAI." CrewAI's absolute count is 2,629,192. AG2's is 215,889.

Do the npm packages tell the same story?

Mastra does, if you pair the package with its own repository. @mastra/core had 11,304,230 downloads and mastra-ai/mastra had 29,512 stars, 383.0 per star. That is a busy package on a repository people actually bookmark. It is an npm count. It is not a PyPI count, and it does not get added to LangGraph's 58,774,245.

The registry documents, read at 05:21:34 GMT and 05:21:35 GMT, are the pairing. @langchain/langgraph's repository field is langchain-ai/langgraphjs, latest 1.7.4. The Python repository langchain-ai/langgraph had 44,934 stars the same morning. 17,998,507 divided by 44,934 is 400.6. That second number uses the wrong repository. The JavaScript package looks "more installed per star" than Mastra because 3,415 is a small denominator, not because a star on langgraphjs means more than a star on Mastra.

The same warning applies to the two SDKs. @openai/agents at 2,195.3 uses openai/openai-agents-js, 4,016 stars. The Python package openai-agents, in the table above, is 625.2 on openai/openai-agents-python, 30,661 stars, and 19,169,518 downloads. They share a product name. They do not share a registry or a star count. @google/adk at 844.6 uses google/adk-js, 1,477 stars. google-adk on PyPI is 480.8 on google/adk-python, 22,209 stars.

Which download counts are dependencies rather than a choice?

langchain-core, pydantic-ai-slim, and the agent-framework meta package. Count them once you know which file pip requested. Do not rank them as a second product on the parent repository's stars.

158,421,619 minus 156,867,095 is 1,554,524. langchain-core was fetched more often than langchain. That fits a file that langgraph also requires, plus any install that asks for core alone. It does not mean core is a more popular framework than LangChain. The current core version at 05:21:13 GMT was 1.8.0, which is the floor of the pin on langchain 1.5.0.

35,124,631 divided by 6,888,240 is 5.10. The slim distribution was fetched about five times as often as the meta package that requires it. Some of that is the meta install. The rest is something else requesting slim, or a cache miss the meta package did not repeat. This note does not split those causes. It refuses to call 1,656.9 "Pydantic AI is the second most installed framework per star." The product row is pydantic-ai at 324.9. The slim row is the dependency.

agent-framework-core had 1,565,472 downloads, 7.87 times the meta package's 198,880. 1,565,472 divided by 198,880 is 7.87. The core row in the first table is the file machines fetched. The meta row is the shorter install line, and it is already inside core's requires. google-adk 2.14.1 lists langgraph only under the all, extensions, and test extras. A default ADK install does not, on that metadata, request langgraph. ADK's 10,678,997 is not an explanation of LangGraph's 58,774,245.

What changed against the previous 30 days?

LlamaIndex's downloads fell. The OpenAI Agents SDK's rose. Stars in the ratio above are one morning, so this section is a download change, not a second per-star series.

The previous window is September 15 through October 14, 2026, taken from the same pypistats responses. It does not overlap October 15 through November 13.

2,901,330 minus 1,936,132 is 965,198. 965,198 divided by 2,901,330 is 0.333, printed as 33.3 percent. The November 5 snapshot had run-llama/llama_index at 52,894 stars. This morning it had 52,893. The download drop is not a star spike in reverse. LlamaIndex's 36.6 downloads per star is a lighter install rate on a still-famous repository, and the latest 30 days are lighter than the 30 days before them.

openai-agents added 6,015,412 downloads, 45.7 percent. smolagents added 249,253, 54.1 percent, and still finished at 23.4 per star because 30,361 stars is a high base under 709,647 downloads. A fast growth rate on a small absolute count does not cross AutoGen's gap, and it does not catch LangGraph. LangGraph added 10,081,233 downloads, 20.7 percent. LangChain fell 9,658,304, 5.8 percent, and remained the largest Python total in the set.

CrewAI's download change was 82,340, 3.2 percent. AutoGen's was 36,926, 9.5 percent. Neither move rearranges the per-star order against LangGraph. The November 13 install note used October 14 through November 12 and a star pull on November 13. Those totals will not match this table. The window is one day later on each end.

What should you do with this if you are choosing a framework?

Read per star when a star sort is about to talk you into a quiet repository, and read the absolute count when a small repository posts a flashy ratio.

If the shortlist is "whatever has the most stars," AutoGen at 6.9 and CrewAI at 43.2 are the rows that should slow you down. LangGraph at 1,308.0, on 58,774,245 downloads, is the orchestration library in this set that people both bookmark and fetch. The Claude Agent SDK and Strands are the rows a star sort misses: under 10,000 stars, over 36 million downloads. Whether either one does the job you have is the framework comparison and the November 14 picker. This page does not award a licence, a checkpointer, or an MCP client.

If you are about to add langchain, langgraph, and langchain-core into one "LangChain installs" number, stop. Core's 158,421,619 already overlaps both. The same rule covers pydantic-ai and pydantic-ai-slim, and agent-framework and agent-framework-core. If you are about to add an npm total to a PyPI total because the product name matches, stop. The registries are different files.

The live series, which moves after this morning's pull, is the downloads tracker and the stars tracker. The dated star-and-release table remains the November 2026 GitHub stats report. A download is still a file request. CI and a transitive dependency are inside the number you just used to make the shortlist shorter.