Top 10 Open-Source Agentic AI Frameworks (2026)
Ranked by GitHub stars, counted through the public API on 27 August 2026 — the one adoption signal in this whole section that can be checked exactly.
Read this first. There is no standardised, comparable usage metric across this market. Vendors publish whichever number flatters them — registered accounts, monthly active users, paying customers, enterprise logos — and those are not interchangeable. Treat the order below as a research shortlist, not a league table.
The ranking
| # | Platform | What it is | Adoption signal | Our review |
|---|---|---|---|---|
| 1 | | Visual builder, API, MCP, multi-agent | 153,724 GitHub stars [github.com] | Not yet reviewed |
| 2 | | Low-code agent platform, RAG, LLMOps | 153,649 GitHub stars [github.com] | Read review |
| 3 | | Agent framework and integration ecosystem | 145,101 GitHub stars [github.com] | Not yet reviewed |
| 4 | | Role-based multi-agent workflows | 57,666 GitHub stars [github.com] | Not yet reviewed |
| 5 | | Data agents, document agents, RAG | 51,888 GitHub stars [github.com] | Not yet reviewed |
| 6 | | Stateful, long-running, durable agents | 40,535 GitHub stars [github.com] | Not yet reviewed |
| 7 | | Multi-agent handoffs, tools, realtime | 29,004 GitHub stars on the Python SDK [github.com] | Not yet reviewed |
| 8 | | Agent SDK for Python, .NET and Java | 28,505 GitHub stars [github.com] | Not yet reviewed |
| 9 | | Production RAG, context engineering, agents | 26,330 GitHub stars [github.com] | Not yet reviewed |
| 10 | | Multi-agent, MCP, Vertex deployment | 21,305 GitHub stars on the Python SDK [github.com] | Not yet reviewed |
The ten in detail
1. Langflow
Top of the list by stars at 153,724, Langflow sits between a prototyping canvas and a runtime you can call over an API. The visual layer earns its keep when non-engineers need to see and edit the graph. The cost is familiar to anyone who has managed generated flow files: diffs get noisy, code review degrades, and debugging drifts from stack traces into a UI. Skip it if your team lives comfortably in code.
2. Dify
Dify is less a library than a self-hosted product: you run a platform, adopt its console, and let a team share prompts and knowledge in one place rather than scattering them across repos. That suits organisations where non-engineers own part of the workflow. It also means operating a stack rather than adding a dependency, and it means living inside someone else’s abstractions the day your requirements stop fitting them. Weigh that before committing.
3. LangChain
One of the most widely adopted options in this category, LangChain offers a shared vocabulary and a broad surface of ready-made connectors, which counts for a lot when you must integrate systems you did not choose., LangChain offers a shared vocabulary and a broad surface of ready-made connectors, which counts for a lot when you must integrate systems you did not choose. The recurring complaint is abstraction: layers that save an afternoon early and cost a week later, when you need to see exactly what reached the model. Sensible for teams willing to read the source; frustrating for teams who expect the framework to stay out of sight.
4. CrewAI
Role-based orchestration is CrewAI’s organising idea: describe agents the way you would describe a small team, assign tasks, let them hand work along. The metaphor makes a design legible in a meeting, which is a real advantage when you must explain the system to people who won’t read code. It can also flatter a design, since delegation between agents tends to add token spend and failure modes, and the hard problems remain control flow and error handling. Prototype before you commit.
5. LlamaIndex
Where most entries here start from orchestration, LlamaIndex starts from the corpus: ingestion, indexing, retrieval, and agents that reason over documents. Pick it when your difficulty is getting the right passages in front of the model, not deciding which tool to call next. If your workload is mostly tool use and long-running control flow, its centre of gravity sits elsewhere, and you’ll end up bolting orchestration on top of a library optimised for data.
6. LangGraph
Sharing a lineage with LangChain but a different philosophy, LangGraph asks you to model the agent as an explicit graph with state you can inspect, persist, and resume. That’s overhead for a chatbot and close to essential for anything that pauses for human approval or runs across hours. The price is conceptual: you think in nodes, edges, and checkpoints before writing much. Teams wanting quick results from a few prompts will find it demanding.
7. OpenAI Agents SDK
Vendor-authored frameworks come with a predictable bargain, and the OpenAI Agents SDK is a clear case: a smaller, more opinionated surface than the general-purpose libraries here, at the cost of a design that follows one company’s roadmap. The licence is open; the gravity isn’t. Reasonable when you have already standardised on that provider and want less framework between you and the model. Worth checking its model-provider options closely when portability across model vendors is a stated requirement.
8. Semantic Kernel
For a shop whose services are written in C# or Java, Semantic Kernel answers a question the rest of this list mostly ignores: what do you build on when Python isn’t your production language? That is the clearest reason to choose it, and for many teams the deciding one. In Python it competes with far more heavily starred libraries, where tutorials, third-party examples, and people who have already hit your bug tend to be easier to find. Language fit should drive this decision, not stars.
9. Haystack
Haystack carries the sensibility of a search-engineering tool rather than an agent demo: pipelines you compose, components you can reason about, and a bias toward what survives production. Its lower star count hints at a smaller community, which can show up as fewer blog posts and fewer copy-paste answers when something breaks — a genuine cost. Worth a look if retrieval quality is the thing your users will judge you on, and if you prefer boring architecture to novel abstractions.
10. Google ADK
At 21,305 stars, Google ADK sits at the bottom of this ranking, which is a popularity signal rather than a verdict on quality. The shape is familiar for cloud-vendor tooling: The shape looks familiar for cloud-vendor tooling: an open framework you can run yourself, with managed Google services close at hand if you want them.. Efficient if your infrastructure already lives with Google, a slow-building liability if it doesn’t, because the low-friction path keeps pointing the same direction. Decide by where you deploy.
How to choose from this list
Read this ranking as a map of attention, not fitness. Two counter-examples make the point: Flowise, with 55,398 stars, was archived on 13 August 2026, and AutoGen, at 60,648 stars, and AutoGen, at 60,648 stars, has taken no commits since 15 April 2026 while its repository remains un-archived.. Both would have placed high here. Stars accumulate; maintenance does not. So sort by your own constraints instead. Ask which part of the problem is actually hard — retrieval, orchestration, or surviving a restart — and which language your production services speak. Then check the commit log and the open issues before you check the star count, and prototype the two finalists on a real task rather than a demo.
Where the figures come from
Every number in the table above was read off a page published by the vendor itself, on 27 August 2026. Figures without a source link were dropped rather than repeated on our own authority.
- Langflow — 153,724 GitHub stars github.com · published counted 27 Aug 2026
- Dify — 153,649 GitHub stars github.com · published counted 27 Aug 2026
- LangChain — 145,101 GitHub stars github.com · published counted 27 Aug 2026
- CrewAI — 57,666 GitHub stars github.com · published counted 27 Aug 2026
- LlamaIndex — 51,888 GitHub stars github.com · published counted 27 Aug 2026
- LangGraph — 40,535 GitHub stars github.com · published counted 27 Aug 2026
- OpenAI Agents SDK — 29,004 GitHub stars on the Python SDK github.com · published counted 27 Aug 2026
- Semantic Kernel — 28,505 GitHub stars github.com · published counted 27 Aug 2026
- Haystack — 26,330 GitHub stars github.com · published counted 27 Aug 2026
- Google ADK — 21,305 GitHub stars on the Python SDK github.com · published counted 27 Aug 2026
Editor's note
Two projects that would place highly on stars alone are deliberately absent. Flowise holds 55,398 stars — enough for fourth place — but its repository was archived on 13 August 2026, so the code is frozen and read-only. Microsoft AutoGen holds 60,648, enough for third, but has taken no commit since 15 April 2026 and its maintainers now point new projects at Microsoft Agent Framework instead; note that GitHub has not flagged the repository as archived, so “maintenance mode” here is the maintainers’ position rather than a platform status. Ranking either on stars would tell you what a project used to be worth, not what it is worth starting with today. And that is the general warning for this table: a star is a bookmark, not a deployment. It records that someone found a repository interesting once, never that anyone still runs it.