On this page

Every list of AI agent frameworks I read quotes a star count, and almost none say when it was taken. A number from last spring sits next to one from last week, and a repo that stopped shipping in April still looks like a contender.

So I built the version I wanted to cite. A script reads 43 open-source agent frameworks from the GitHub, npm and PyPI APIs every day and rebuilds the table below. Every figure carries its date, the scoring is written out in full, and the data is free to copy.

This page is the dataset. If you want my opinion on which framework to pick for a project, that lives in the best open-source AI agent frameworks, which compares the six I would shortlist and explains why.

Headline numbers
  1. Dify ranks first of 43 open-source agent frameworks on the composite score (96.3 of 100) as of 4 October 2026, ahead of n8n (93.1) and Mastra (84.2). Among code-first frameworks, Mastra leads at 84.2, then LangChain at 84.0.
  2. Dify gained the most GitHub stars: about 3,446 in the last 30 days, followed by n8n (about 3,356) and AgentScope (about 2,426). Relative to size, AgentScope grew fastest among repos above 5,000 stars, at 8.0% in 30 days.
  3. Mastra was the most active repo, with 1,662 commits to its default branch in 30 days; n8n (1,402) and Pipecat (1,065) follow.
  4. 32 of the 43 frameworks shipped a release in the 30 days to 4 October 2026, while 6 are archived or had no commit in 90 days (AutoGen, MetaGPT, Flowise, SuperAGI, Upsonic, Inngest AgentKit).
  5. The 43 tracked repos hold 1,748,887 GitHub stars between them; the visual builders (Dify, n8n, AutoGPT, Langflow, Activepieces, Flowise, Coze Studio) account for 46% of them.

Updated . Refreshed daily from the GitHub, npm and PyPI APIs.

Cite this page: "Open-source AI agent framework leaderboard", Gravity, updated , https://gravity.fast/data/ai-agent-framework-leaderboard/. Data licensed CC BY 4.0: reuse it anywhere, commercially too, with credit and a link. Download the full table as data.csv.

The leaderboard

Ranked by the composite score described in the methodology. The Type column separates code-first libraries from visual builders and voice-agent frameworks, so sort by eye within a type if that is the comparison you need. Repo names link to GitHub.

#FrameworkTypeStarsStars gained, 30 daysCommits, 30 daysActive contributors, 90 daysLatest releaseOpen issuesWeekly downloadsScore
1DifyVisual builder, TypeScript157,824+3,446*77221210 September 20262%No package96.3
2n8nVisual builder, TypeScript206,632+3,356*1,4021712 October 20263%137.6k npm93.1
3MastraCode framework, TypeScript28,551n/a1,66221430 September 20265%2.2M npm84.2
4LangChainCode framework, Python147,433+1,787*192622 October 20265%40.7M PyPI84.0
5OpenAI Agents SDK (Python)Code framework, Python29,829n/a2551122 October 20260%3.0M PyPI77.2
6AutoGPTVisual builder, Python187,651+518*4974630 September 20268%No package77.0
7Google ADKCode framework, Python21,702n/a4582182 October 20267%2.3M PyPI76.5
8CrewAICode framework, Python59,339+1,320*1075528 September 20269%606.0k PyPI73.9
9Vercel AI SDKCode framework, TypeScript27,116+546*502781 October 202610%34.2M npm72.4
10LangflowVisual builder, Python155,500+1,208*1514529 September 20267%13.8k PyPI72.1
11AgentScopeCode framework, Python32,751+2,426*1509228 September 202627%26.5k PyPI71.9
12Pydantic AICode framework, Python20,398+680*7521303 October 202626%1.4M PyPI71.8
13LangGraphCode framework, Python42,704+1,721*561623 September 202635%11.5M PyPI68.6
14PipecatVoice agents, Python16,172n/a1,0659126 September 20268%257.0k PyPI68.3
15LiveKit AgentsVoice agents, Python14,486+764*1871591 October 202611%976.5k PyPI66.7
16AgnoCode framework, Python42,547+511*1201232 October 202627%500.9k PyPI66.4
17DSPyCode framework, Python38,500+725*603725 September 202621%1.4M PyPI66.0
18Microsoft Agent FrameworkCode framework, Python13,938n/a4921552 October 202615%56.1k PyPI64.9
19Strands AgentsCode framework, Python8,651n/a2481151 October 202635%8.7M PyPI64.8
20ActivepiecesVisual builder, TypeScript24,894+650*3953030 September 202614%No package64.1
21HaystackCode framework, Python26,649+236*2551231 October 20262%161.3k PyPI62.3
22LlamaIndexCode framework, Python52,407+389*255421 September 20263%687.3k PyPI58.8
23OpenAI Agents SDK (JS)Code framework, TypeScript3,887+136*1572710 September 20261%2.4M npm43.4
24SwarmsCode framework, Python7,230n/a105123 October 2026 (PyPI)14%7,738 PyPI39.8
25LettaCode framework, n/a25,027n/a212 October 2026 (PyPI)0%12.6k PyPI39.1
26Semantic KernelCode framework, C#28,625+104*14143 September 20262%80.1k PyPI38.5
27smolagentsCode framework, Python29,669+532*4129 May 202639%106.7k PyPI37.2
28BeeAI FrameworkCode framework, Python3,426n/a583228 September 20262%8,793 PyPI36.3
29AG2Code framework, Python4,977+79*102263 October 20263%33.8k PyPI35.7
30AutoGenCode framework, Python61,254+431*0030 September 202518%88.7k PyPI35.6
31LangGraph.jsCode framework, TypeScript3,335+77*60243 October 202621%4.5M npm35.6
32MetaGPTCode framework, Python70,739+549*009 March 20251%702 PyPI35.4
33VoltAgentCode framework, TypeScript10,724+226*2728 September 202614%50.8k npm33.5
34CAMELCode framework, Python17,811+142*13213 September 202616%16.3k PyPI33.4
35LangroidCode framework, Python4,111n/a29102 October 202617%9,572 PyPI32.5
36Agent SquadCode framework, Python7,782n/a15423 September 202617%2,233 PyPI31.4
37Atomic AgentsCode framework, Python6,268n/a121127 September 20263%3,975 PyPI30.8
38Flowise archivedVisual builder, TypeScript55,489+66*0429 July 202626%2,785 npm30.5
39Agency SwarmCode framework, Python4,591n/a7163 August 20262%2,116 PyPI26.7
40Coze StudioVisual builder, TypeScript21,674n/a015 February 202647%No package21.3
41SuperAGICode framework, Python17,698n/a0016 January 202440%No package16.7
42UpsonicCode framework, Python7,956n/a0019 May 202610%5,046 PyPI11.8
43Inngest AgentKitCode framework, TypeScript940n/a0013 November 202553%32.7k npm7.1

Data as of 4 October 2026. Stars, commits and active contributors (distinct commit authors on the default branch in 90 days, bots excluded) and releases from the GitHub API; weekly downloads are the last 7 days on npm (api.npmjs.org) or PyPI (ClickPy's copy of the PyPI download logs, with pypistats.org as fallback) for the package named in each project's install docs. "Open issues" is open issues as a share of all issues ever filed, pull requests excluded. * Stars gained is scaled to 30 days from a Wayback Machine capture of the repo page between 20 and 75 days old; it switches to this tracker's own daily snapshots once they are 30 days deep. Score: see methodology.

Who is gaining stars

Total stars reward age. A framework launched in 2023 has had three years to collect them, so the more useful question for anyone choosing today is who is gaining attention now. The chart ranks the repos by stars gained in the last 30 days.

GitHub stars gained in the last 30 days, top 12 open-source agent frameworks Dify 3,446; n8n 3,356; AgentScope 2,426; LangChain 1,787; LangGraph 1,721; CrewAI 1,320; Langflow 1,208; LiveKit Agents 764; DSPy 725; Pydantic AI 680; Activepieces 650; MetaGPT 549. Data as of 4 October 2026. STARS GAINED IN 30 DAYS, AS OF 4 OCTOBER 2026 Dify ~3,446 n8n ~3,356 AgentScope ~2,426 LangChain ~1,787 LangGraph ~1,721 CrewAI ~1,320 Langflow ~1,208 LiveKit Agents ~764 DSPy ~725 Pydantic AI ~680 Activepieces ~650 MetaGPT ~549
Stars gained in the 30 days to 4 October 2026, from the GitHub API. Values marked ~ are scaled from a Wayback Machine capture of the repo page, as the table note explains.

Absolute gains favour big repos, so the headline numbers also name the fastest grower relative to its size among repos above 5,000 stars. Both are in the CSV if you want to run your own cut.

How to read the ranking

Stars are attention, not usage. A star costs one click and is never taken back by most people. It tells you how many developers noticed a project, which is why the visual builders, with their broad non-developer audiences, hold most of the stars in this table.

Downloads are installs, inflated by machines. Weekly npm and PyPI counts include CI pipelines, container builds and other libraries that depend on a package. LangChain and the Vercel AI SDK post tens of millions a week partly because so much other software pulls them in. Compare downloads within a registry, and treat them as a measure of how embedded a package is.

Commits and active contributors are the pulse. A repo with 60,000 stars and no commits in 90 days is a monument. The table shows that directly: as of 4 October 2026, AutoGen's default branch had no commits in the previous 90 days while Microsoft Agent Framework had hundreds, and Flowise's maintainers had archived the repo on GitHub. If you run Flowise today, these are the alternatives I compared.

The score is one opinion made explicit. I weighted momentum (growth, commits, contributors) at 55% and accumulated popularity (stars, downloads) at 35%, because a framework you adopt today needs to be maintained next year. If you weight things differently, the CSV has every input.

A note on where I sit. I run Gravity, which is on the other side of this choice: people describe a task and an agent runs it, so nobody using it picks a framework. That gives me no favourite in this table, and it is why the comparisons I write, like Gravity vs LangGraph and Gravity vs CrewAI, are about build-it-yourself versus hand-it-off rather than one framework versus another.

Change log

Each line is generated from the daily data: releases published in the last week, rank moves of three places or more, star milestones, and repos archived by their maintainers.

  1. : Leaderboard launched with 43 open-source agent frameworks, refreshed daily.
  2. : Pydantic AI published release v2.54.0.
  3. : AG2 published release v1.1.2.
  4. : LangGraph.js published release @langchain/langgraph-cli@1.5.2.
  5. : n8n published release n8n@2.41.6.
  6. : LangChain published release langchain-text-splitters==1.1.3.
  7. : OpenAI Agents SDK (Python) published release v0.23.1.
  8. : Google ADK published release v2.11.0.
  9. : Agno published release v3.1.1.
  10. : Microsoft Agent Framework published release python-1.20.0.
  11. : Langroid published release 0.68.4.
  12. : Vercel AI SDK published release @ai-sdk/harness-github-copilot@1.0.34.
  13. : LiveKit Agents published release livekit-agents@1.8.4.
  14. : Strands Agents published release python/v1.57.2.
  15. : Haystack published release v3.3.0.
  16. : Mastra published release @mastra/core@1.72.0.
  17. : AutoGPT published release autogpt-platform-beta-v0.8.2.
  18. : Activepieces published release 0.92.1.
  19. : Langflow published release v1.12.4.
  20. : CrewAI published release 1.15.23.
  21. : AgentScope published release v2.0.9.
  22. : BeeAI Framework published release python_v0.1.85.
  23. : VoltAgent published release @voltagent/core@2.11.0.
  24. : Atomic Agents published release v2.10.3.

Methodology

What gets included

A project is in the table when it is open source on GitHub, its main purpose is building LLM agents (tool calling, multi-step plans, multi-agent coordination, or a visual canvas for the same), and it has meaningful public use. I confirmed every repo through the GitHub API on 4 October 2026 and recorded renames, such as Strands Agents moving to strands-agents/harness-sdk and Atomic Agents to Eigenwise/atomic-agents.

Left out on purpose: coding agents that are products rather than frameworks (OpenHands, Goose), memory and retrieval layers used inside agents, model SDKs with no agent loop, and ports that were archived and folded into another repo. Where a project ships separate Python and JavaScript repos with their own communities, as LangGraph and the OpenAI Agents SDK do, each repo gets its own row.

The seven measures

  • Stars: the repo's stargazer count from the GitHub GraphQL API, read daily.
  • Stars gained, 30 days: today's count minus the count 30 days earlier from this tracker's own snapshots. Until those exist, the baseline is the nearest Wayback Machine capture of the repo page taken 20 to 75 days earlier, scaled to 30 days and marked with an asterisk. Repos with no usable capture show n/a. GitHub no longer serves the list of who starred a repo and when, so star history cannot be rebuilt from the API.
  • Commits, 30 days: commits reachable from the default branch with a commit date in the last 30 days. Squash-merge projects show fewer commits for the same amount of work, so read this alongside contributors.
  • Active contributors, 90 days: distinct commit authors on the default branch in the last 90 days, counted by GitHub login or commit email, with bots such as Dependabot and Renovate removed. I use this rather than GitHub's all-time contributor list because that list stops linking accounts after about 500 authors and cannot rank large repos.
  • Latest release: the newest GitHub release that is not a draft or pre-release, or the newest upload of the project's PyPI package when that is more recent (marked PyPI).
  • Open issues: open issues as a share of all issues ever filed, pull requests excluded. A low share means issues get closed; very young repos and repos that use another tracker read oddly here.
  • Weekly downloads: the last complete 7 days for the package named in the project's install docs, from the npm downloads API or from ClickPy, ClickHouse's public copy of the PyPI download logs, with pypistats.org as the fallback.

The composite score

For each measure, every framework gets its percentile rank within the tracked set, from 0 for the lowest to 1 for the highest. The score is the weighted average of those ranks, times 100:

MeasureWeightWhy
Stars25%Accumulated attention; the number most people quote
Stars gained, 30 days20%Attention now rather than in 2023
Commits, 30 days20%Is anyone shipping
Active contributors, 90 days15%How many people the project depends on
Weekly downloads10%How embedded the package is
Release recency10%Full weight for a release in 30 days, half for one in 90 days, none after

Missing measures (no package to download, no star baseline yet) are dropped and the remaining weights rescaled, so a project is never penalised for data that does not exist. Ties break on stars.

Known gaps

  • Stars and downloads can be gamed, and neither is filtered for that here.
  • Download counts mix registries: an npm download and a PyPI download are not the same unit, which is why downloads carry only 10% of the score.
  • Projects that develop in a private repo and mirror to GitHub show fewer commits and contributors than they have.
  • The 30-day growth baseline comes from archived pages until early November 2026, when this tracker's own snapshots take over.

The refresh runs daily. The code that builds this page reads only public APIs, and corrections are welcome at support@gravity.fast. For what these frameworks and platforms cost once you run them in production, see the AI agent pricing index and the cheapest AI agent platforms. The same collector also powers the MCP and A2A adoption census.

FAQ

What is the most popular open-source AI agent framework?

It depends on the measure. By GitHub stars on 4 October 2026, n8n (206,631) and AutoGPT (187,651) lead, followed by Dify, Langflow and LangChain, all above 145,000. By weekly package downloads, LangChain leads on PyPI (about 40.7 million in the week to 3 October 2026) and the Vercel AI SDK on npm (about 34.2 million). By the composite score on this page, which also weighs recent growth, commits, active contributors and releases, the order changes daily, so check the table for today's ranking.

Which AI agent framework is growing fastest?

The headline numbers at the top of the page name the repo that gained the most GitHub stars in the last 30 days, and the chart shows the top 12. Until this tracker's own snapshots are 30 days deep, growth is scaled from a Wayback Machine capture of each repo page taken 20 to 75 days earlier, and those values carry an asterisk.

LangGraph vs CrewAI: which is more popular?

CrewAI has more GitHub stars (59,339 against 42,704 on 4 October 2026), while LangGraph has far more package downloads: about 11.5 million a week on PyPI against about 0.6 million for CrewAI in the week to 3 October 2026. Stars measure attention; downloads measure how often the code is installed, including by other libraries and CI systems.

How is the composite score calculated?

Each framework gets a percentile rank among the tracked set on five measures: stars (25%), stars gained in 30 days (20%), commits in 30 days (20%), active contributors in 90 days (15%) and weekly downloads (10%). A release in the last 30 days adds the full 10% release weight, one in the last 90 days half of it. When a measure is missing, such as downloads for a project with no package, the remaining weights are rescaled. The result is shown out of 100.

Why are n8n, Dify and Langflow on an agent framework leaderboard?

Because many teams build agents with them instead of with a code library, and searchers comparing open-source agent options meet them in the same results. They are labelled Visual builder in the Type column so you can compare like with like.

How often is the leaderboard updated, and can I reuse the data?

A script refreshes every number daily from the GitHub, npm and PyPI APIs. The data is published under CC BY 4.0: copy the table or download the CSV, and credit Gravity with a link to this page.

Why is my favourite framework missing?

The list covers open-source projects on GitHub that developers use to build LLM agents, with a public repo and meaningful usage. Coding agents such as OpenHands and Goose are products rather than frameworks and are left out. To suggest a framework, email support@gravity.fast with the repo URL; additions show up in the change log.

Sources

  1. GitHub GraphQL API: stars, issues, releases, commit history and authors for each repo, read daily from 4 October 2026.
  2. npm download counts API: last-week downloads for npm packages, read daily.
  3. ClickPy: ClickHouse's public dataset of PyPI downloads, built from the PyPI BigQuery logs, read daily.
  4. pypistats.org API: fallback for PyPI weekly downloads.
  5. PyPI JSON API: latest package version and upload date.
  6. Wayback Machine CDX API: archived GitHub repo pages from August and September 2026, used once for 30-day star baselines.