Best Open-Source AI Agent Frameworks in 2026: 15 Frameworks for Teams That Want to Own Their Stack

The best open-source AI agent framework in 2026 depends more on your licence tolerance and your runtime than on any feature checklist: LangGraph and CrewAI lead production Python deployments, Mastra and Flowise lead if your team lives in TypeScript, and Dify covers the widest surface if you want a self-hosted app instead of a code library. This guide ranks 15 frameworks that are genuinely open source (a public repo, a real licence, self-hostable on your own infrastructure), not just free to try, and it names the handful that ship under an open-core or source-available licence instead of a pure permissive one.

Picking wrong here costs more than picking the wrong SaaS tool. You're choosing an abstraction your engineers will write thousands of lines against, and migrating off a graph-based orchestration layer six months in is a multi-week project, not a re-export. Every entry below is evaluated against its actual LICENSE file, current GitHub activity, language support, state and durability model, and hosted pricing where one exists, all verified against the vendor's own pages and repos in August 2026. Updated August 2026. If you're still deciding between a self-hosted framework and a managed product, our best AI agent platforms roundup covers the full landscape both sides live in.

Key Facts

• More than 1.1 million public GitHub repositories now import an LLM SDK, up 178% year over year, with roughly 694,000 of those created in the past 12 months alone, per GitHub's 2025 Octoverse report.
• 57% of organizations report having AI agents running in production, per LangChain's State of Agent Engineering survey of 1,340 practitioners (November 18 to December 2, 2025), though only 52.4% run offline evaluations against that traffic.
• Gartner predicts more than 40% of agentic AI projects will be canceled by the end of 2027 over unclear ROI and weak governance, and estimates only about 130 of the thousands of self-described agentic AI vendors are the real thing.
• Open-source models' share of enterprise LLM usage fell to 11% in 2025, down from 19% in 2024, even as total enterprise generative AI spending roughly tripled to $37 billion, per Menlo Ventures.
• AI agents jumped from roughly 12% to 66.3% task success on OSWorld, the benchmark for real computer-use tasks, in the year leading into 2026, per Stanford HAI's AI Index.

What Changed in Open-Source Agent Frameworks This Year

• Microsoft merged AutoGen and Semantic Kernel into Microsoft Agent Framework, which reached general availability on April 3, 2026. Both legacy repos are now officially in maintenance mode.
• AG2, the community fork of AutoGen created after a 2024 governance split, kept building independently under Apache-2.0 and is working toward a 1.0 release.
• OpenAI is sunsetting the Assistants API on August 26, 2026. If you're building new, use the Agents SDK covered here, not the older Assistants API.
• CrewAI simplified its commercial pricing down to two public tiers (free AMP Basic and custom Enterprise) after briefly running a self-serve Professional plan.
• IBM closed its acquisition of DataStax in May 2025, which means Langflow now sits inside IBM's watsonx orbit.
• Workday acquired FlowiseAI in August 2025, folding the visual builder into its enterprise HR and finance agent roadmap.

Quick Comparison Table

Use this to shortlist by licence and language before reading the licence and self-hosting detail below.

How to Read "Open Source" Here: Licence Reality Check

Not everything billed as "open source" in the agent space carries the same rights, and this is the detail most roundups skip. An OSI-approved licence (MIT, Apache-2.0) puts no restriction on how you use, modify, or resell the code. A source-available or open-core licence publishes the code but reserves a specific right, usually the right to resell it as your own hosted SaaS, for the vendor. That distinction rarely shows up until legal review, which is too late.

Dify is the clearest case: its licence is a modified Apache-2.0 (the "Dify Open Source License") that explicitly blocks running Dify to operate a multi-tenant SaaS without a commercial licence from LangGenius, and it locks the console's branding outside a paid tier. That's a real, meaningful restriction an OSI licence would not permit, which makes Dify source-available rather than open source in the strict sense, even though the vast majority of its code is genuinely Apache-2.0. Mastra and Flowise both take a cleaner open-core approach: their core frameworks are unrestricted Apache-2.0, and only a specific enterprise directory (ee/ for Mastra, enterprise/ for Flowise) sits behind a separate commercial licence. Everything you'd need to self-host a production agent in either framework is in the open part.

None of this stops a company from self-hosting any of the 15 for internal use. It matters the moment you plan to embed one inside a product you sell.

Self-Hosting, Language, and State: The Engineering Fit

The 15 frameworks split into three real categories: code-first orchestration libraries you import into a service (LangGraph, CrewAI, AutoGen/AG2, OpenAI Agents SDK, Pydantic AI, Semantic Kernel, LlamaIndex Workflows, Agno, Mastra, smolagents, Haystack), self-hosted visual or full-app platforms (Dify, Langflow, Flowise), and a specialized coding-agent runtime (OpenHands). All 15 run entirely on your own infrastructure with no forced call-home. What differs is how much durability you get for free versus how much you build yourself.

If your agent needs to survive a server restart mid-task, that "built-in" column is the one to read closely, not the star count.

Maintenance and Community Health: Who's Actually Still Building

A star count alone tells you almost nothing about whether a project is safe to build on. Governance and recent activity tell you much more: a company-controlled repo can enter maintenance mode overnight, the way AutoGen and Semantic Kernel just did, while a smaller, community-governed fork can keep shipping past the point where the company that started it moves on.

Stars and push dates are a snapshot pulled directly from each repository's GitHub API record on August 17, 2026, not a permanent ranking, they change daily. The pattern worth noting: 14 of these 15 projects are led by a single company, which means the roadmap, and occasionally the licence, can shift whenever that company's strategy does. AG2 is the one project here with genuinely open, non-corporate governance.

1. LangGraph: The Production Standard for Stateful Agents

LangGraph, from LangChain Inc, models an agent as a graph of nodes and edges instead of a single prompt loop, which is what makes pause, resume, branching, and "time travel" debugging native rather than bolted on. That checkpointing model is the main reason it shows up so often as the production default once a team moves past a prototype. It works standalone or alongside the broader LangChain ecosystem, and ships for both Python and TypeScript.

The trade-off is the graph mental model itself: it takes longer to learn than a role-based crew, and the best observability layer, LangSmith, is free for one seat but priced per seat beyond that. For a hands-on walkthrough, see our build an AI agent with LangGraph tutorial.

Licence: MIT. Best for: Teams building agents that must pause, resume, and recover state without losing progress.

2. CrewAI: Role-Based Crews, Fastest Time to a Working Demo

CrewAI defines agents by role, goal, and backstory, groups them into a crew, and runs them sequentially or hierarchically against a set of tasks. That structure, plus a large tutorial and course ecosystem, is why it's often the fastest path from zero to a working multi-agent demo. Flows add a more deterministic, code-first control layer for teams who outgrow pure autonomous crews.

CrewAI Inc's hosted AMP platform handles deployment and observability, and pricing simplified in 2026 to a free Basic tier (50 executions per month) and a custom Enterprise tier, after a self-serve Professional plan was discontinued. For hands-on setup, see build an AI agent with CrewAI.

Licence: MIT. Best for: Teams that want role-based multi-agent collaboration running in an afternoon.

3. AutoGen and AG2: One Framework, Two Paths After the Split

Microsoft's AutoGen was one of the first widely-adopted multi-agent conversation frameworks. In late 2024, several of its original creators left Microsoft and forked the project into AG2, first as a GitHub fork and then as its own organization, inheriting the autogen and pyautogen PyPI package names in the process. The split was about governance: Microsoft kept AutoGen company-led, while AG2 became community-governed.

As of 2026, Microsoft's AutoGen is officially in maintenance mode: no new features, with development redirected into Microsoft Agent Framework (merged with Semantic Kernel, GA April 3, 2026). AG2, by contrast, is actively developed under Apache-2.0, with a published roadmap toward 1.0 and cross-framework interoperability work under an "AgentOS" banner. If you want the conversational multi-agent pattern AutoGen popularized, with ongoing open-source development, AG2 is now the living project. For the underlying pattern either way, see multi-agent systems.

Licence: AutoGen: MIT (code), CC-BY-4.0 (non-code assets). AG2: Apache-2.0. Best for: Teams already invested in AutoGen's pattern who need to decide between following Microsoft or staying open source with AG2.

4. OpenAI Agents SDK: The Lightest SDK, Built by the Model Vendor

The OpenAI Agents SDK is the production successor to OpenAI's experimental "Swarm" pattern: a deliberately small set of primitives (Agents, Handoffs, Guardrails, Sessions) built on top of the Responses API. Despite the name, it isn't locked to OpenAI models. An official LiteLLM extension gives it access to more than 100 providers, including Anthropic, Gemini, and Bedrock, through the same agent code.

Worth stating plainly: this is a different product from the OpenAI Assistants API, which OpenAI is retiring on August 26, 2026. The Agents SDK is the actively developed path forward; the Assistants API is not.

Licence: MIT. Best for: Teams that want the smallest possible dependency footprint and are comfortable wiring their own persistence.

5. Pydantic AI: Type-Safe Agents for FastAPI-Trained Teams

Pydantic AI comes from the team behind Pydantic itself, the validation library most Python web APIs already depend on, and it shows: agent outputs are structured and validated by construction rather than parsed hopefully from a text blob, and tool dependencies are injected the way FastAPI injects request dependencies. That makes malformed tool calls a type error you catch in testing, not an incident you catch in production.

The framework itself is fully open and free. Logfire, the team's optional observability product, is separate: a free tier with 10 million spans a month, a Team tier at $49 a month for five seats, and a Growth tier at $249 a month for unlimited seats, all confirmed on pydantic.dev/pricing.

Licence: MIT. Best for: Python teams that want agent outputs to fail a type check instead of failing in production.

6. Microsoft Semantic Kernel: Enterprise .NET and Python, Now in Maintenance Mode

Semantic Kernel built its reputation on enterprise connectors: Azure AD, Key Vault, Application Insights, and a plugin/planner model that mirrors how .NET shops already structure business logic. It's also the only framework on this list with mature Java support alongside .NET and Python.

The fact that matters most for anyone evaluating it new: Microsoft merged Semantic Kernel with AutoGen into Microsoft Agent Framework, generally available since April 3, 2026, and Semantic Kernel itself is now in maintenance mode. It still receives security fixes, but new orchestration patterns are landing in the successor, not here.

Licence: MIT. Best for: Teams with an existing Semantic Kernel investment who need a stable bridge while planning a Microsoft Agent Framework migration.

7. LlamaIndex Workflows: Event-Driven Orchestration From a RAG Pedigree

LlamaIndex Workflows models an agent as a set of async steps that emit and consume typed events, so a multi-step agent reads like a small state machine instead of a dedicated graph DSL. It grew directly out of LlamaIndex's data-connector and retrieval heritage, LlamaHub included, which shows in how naturally it handles agents whose core job is pulling from documents and structured data rather than pure conversation.

LlamaIndex Inc's hosted LlamaCloud handles parsing and indexing at usage-based rates, kept separate from the orchestration framework, which stays free and open.

Licence: MIT. Best for: Teams building agents where retrieval over documents or data is the core job, not an add-on.

8. Agno: The Performance-First Framework Formerly Known as Phidata

Agno, renamed from Phidata in January 2025, now positions itself less as a library and more as an "agentic operating system," built around Agents, Teams, and Workflows abstractions plus a runtime called AgentOS that owns durable state and exposes it over HTTP. The pitch is speed: Agno's own published benchmarks claim roughly 5,000 times faster agent instantiation and about 50 times less memory than LangGraph. That's a vendor claim worth benchmarking against your own workload before you rely on it, not an independently verified figure.

Licence: Apache-2.0. Best for: Teams running large numbers of concurrent lightweight agents where instantiation overhead is a real, measured bottleneck.

9. Mastra: TypeScript-Native Agents for Full-Stack JS Teams

Mastra, built by a team that includes former Gatsby engineers, is the first framework on this list designed API-first for TypeScript rather than ported from a Python design. Workflows, agents, RAG, evals, and a memory layer are built to sit naturally inside an existing Next.js or Node service, which matters if standing up a separate Python microservice just to run agents isn't worth the operational overhead.

The core framework is Apache-2.0. An ee/ directory carries a separate Mastra Enterprise licence, source-available for development but requiring a paid licence in production, and the optional Mastra Cloud control plane starts free (100,000 observability events, 24 CPU hours, 15-day retention) with a Teams tier at $250 a month, confirmed on mastra.ai/pricing.

Licence: Apache-2.0 (core), separate commercial licence for ee/. Best for: JavaScript and TypeScript teams that don't want to stand up a separate Python service just to run agents.

10. smolagents: Hugging Face's Minimal, Code-Writing Agents

smolagents takes a different bet than the rest of this list: instead of chaining JSON tool calls, its core "CodeAgent" writes and executes small pieces of Python as its primary action. Hugging Face's argument is that this cuts the number of steps and the error rate compared to classic tool-calling loops, since a model writing results = [search(q) for q in queries] does in one step what three sequential tool calls would need. The core library is deliberately small, small enough to read end to end in an afternoon.

Licence: Apache-2.0. Best for: Teams that want to understand every line of their agent loop, or that are already deep in the Hugging Face ecosystem.

11. Haystack: Production Retrieval Pipelines With Agents Built In

Haystack, maintained by the German company deepset, started as a search and retrieval pipeline framework years before "agent" became the industry's word of the month, and that heritage shows in how explicit and battle-tested its component and pipeline model is for retrieval and document processing. Agents in Haystack are built as components inside the same pipeline graph, not a separate bolt-on concept.

deepset sells a Haystack Enterprise Starter (support plus production templates) and a full Haystack Enterprise Platform on top of the open core, both priced by organization size rather than published as a rate card.

Licence: Apache-2.0. Best for: Teams whose agent's real job is accurate retrieval and document handling, with conversation as a thin layer on top.

12. Dify: The Full LLMOps App, With a Licence You Should Actually Read

Dify packages a visual workflow builder, a prompt IDE, a RAG pipeline, agent orchestration, and observability into one self-hosted application, which makes it closer to a full LLMOps platform than a code library. That breadth, plus a large community, is likely why it sits among the most-starred repositories in this entire list, second only to Langflow. LangGenius, the company behind it, also runs Dify Cloud: a free Sandbox tier, Professional at $59 a month ($590 a year billed annually), Team at $159 a month ($1,590 a year billed annually), and a custom Enterprise tier, all confirmed on dify.ai/pricing.

Read the licence section above before betting a resold product on this one: Dify's modified Apache-2.0 licence blocks operating it as a multi-tenant SaaS without a commercial licence from LangGenius. For a fully-managed alternative that sidesteps the self-hosting question entirely, see our best no-code AI agent builders roundup.

Licence: Modified Apache-2.0 ("Dify Open Source License"), source-available. Best for: Teams that want an internal, self-hosted LLM app builder for employees, not a codebase to embed inside a resold product.

13. Langflow: Drag-and-Drop Building, Now Inside IBM's Orbit

Langflow puts a visual canvas over LangChain-style components, and it exports what you build as runnable Python or a callable API endpoint, which makes it a genuinely popular way to prototype an agent flow and hand it to engineers, or to demo one to a non-technical stakeholder. DataStax acquired Langflow in April 2024, and IBM closed its acquisition of DataStax in May 2025, so Langflow's roadmap now sits inside IBM's watsonx strategy.

Unlike Dify or Flowise, Langflow carries no licence carve-out: it's MIT, full stop.

Licence: MIT. Best for: Teams that want a free, MIT-licensed visual builder for prototyping before committing a flow to hand-written code.

14. Flowise: The JavaScript-Native Visual Builder Workday Just Bought

Flowise is the TypeScript-native answer to Langflow's Python-first canvas, which matters if your backend is already a Node service and you don't want to bring Python into the stack just for agent prototyping. Workday announced its acquisition of FlowiseAI in August 2025, confirmed in Workday's own newsroom, and the tool is now expected to feed Workday's enterprise HR and finance agent roadmap over time.

The core framework stays Apache-2.0, with an enterprise/ directory carrying a separate proprietary licence. Flowise Cloud runs free (2 flows, 100 predictions a month), Starter at $35 a month (10,000 predictions), and Pro at $65 a month (50,000 predictions, five users included, $15 per extra user), all confirmed directly on flowiseai.com.

Licence: Apache-2.0 (core), proprietary licence for enterprise/. Best for: JavaScript teams that want the Langflow experience without bringing Python into the stack.

15. OpenHands: The Open-Source Autonomous Software Engineer

OpenHands, renamed from OpenDevin in late 2024 and built by All Hands AI, does a different job than the other 14 frameworks on this list. It's not a general-purpose library you assemble a custom agent on top of, it's a specialized runtime that gives an agent a real sandboxed shell, browser, and code editor so it can plan, write, run, test, and fix code autonomously. If you're comparing it against hosted coding-agent products instead, see our best AI coding agents roundup.

All Hands AI runs OpenHands Cloud alongside the free local version: an Individual tier that's free with your own API key or pay-at-cost model access, and a custom Enterprise tier for private-VPC or self-hosted deployment with SSO, confirmed on openhands.dev/pricing.

Licence: MIT. Best for: Teams that want an open, self-hostable alternative to closed coding-agent products, not a framework to build a general-purpose agent on.

The Real Cost of "Free"

Every framework's code is free in all 15 cases. What isn't free is everything that makes an agent trustworthy enough to run unattended. Almost every vendor on this list monetizes the same layer: not the orchestration code, but the observability, evaluation, and managed hosting wrapped around it.

Budget for that last row before you budget for a licence fee. A framework with a $0 price tag and no engineer who understands its state model is more expensive than a framework with a modest hosted tier and a team that's fluent in it.

How to Choose: Decision Framework

Match your situation to the right starting point, then read that framework's full section above before committing.

If your team would rather evaluate a fully managed platform instead of assembling one from a framework, our best enterprise AI agent platforms roundup and our choosing an AI agent platform guide both cover that path.

What to Do Next

Don't evaluate all 15 frameworks at once. Narrow to two based on your language (Python or TypeScript) and your build style (code library or visual app), then build the same small agent, one tool call, one multi-step handoff, and one deliberate failure case, in both. The framework that survives you trying to break it is the one to build on.

If your team would rather not build the agent at all, our best AI agent platforms roundup covers managed alternatives, and the how to build an AI agent guide walks through the underlying steps regardless of which framework you end up choosing.