Open-Source AI Agent Frameworks
Open-source frameworks for building AI agents on code you can read, fork and self-host: from the largest ecosystems to the newest TypeScript-native entrant.
On this page
12 of the 14 agent frameworks in our index carry a clean, unrestricted open-source license on the code you would actually run. No source-available carve-out. No separate commercial license required to get the controls a production team needs while self-hosting. A framework is not a product you sign into; it is code that lands in your own repository, so the license attached to it matters more here than it does for a hosted agent.
Two of the 14 do not clear that bar, and both are open at the core. Mastra's framework is Apache-2.0 and free to self-host, but the RBAC, SSO and IAM controls a production team needs sit behind a separate Enterprise Self-Hosted tier, billed as a flat annual fee on top of the open code. Claude Agent SDK's Python and TypeScript wrapper packages carry an MIT license, but Anthropic's own docs state that using the SDK as a whole is governed by its commercial Terms of Service, so the license on the wrapper does not cover what you are actually agreeing to when you run it.
This is not the same list as our [open-source AI agent tools collection](/collections/open-source-ai-agent-tools/), which spans coding agents, voice agents and agent platforms too, not just the tools you build an agent with (every entry here also appears there). It is also not the same as our [ranking of all 14 frameworks](/best-ai-agent-frameworks/), which scores the full set on capability and includes Mastra and Claude Agent SDK alongside the 12 open ones here.
Among these 12: LangChain remains the largest and most integration-heavy framework in the group; Microsoft ships two that clear this bar, AutoGen and its newer successor, Agent Framework; and AgentsKit is the newest and least proven, four months old with an MIT core and its hosted layer, AgentsKit OS, still enterprise-only early access. Every one of them lets you point the framework at any model provider you choose and keep running if the vendor ever disappears, a guarantee a closed SDK cannot make.
Why this collection
A framework earns a place here only if it belongs to this site's frameworks category and its entire self-hostable codebase carries a clean, unrestricted OSI-style license: MIT, Apache-2.0, GPL-3.0-or-later or AGPLv3, with no source-available carve-out and no separate commercial license gating the controls a production deployment needs. This is the same bar our open-source AI agent tools collection applies, narrowed to one category. Verified against each framework's own listing research (license type, pricing tiers and FAQ text), not a vendor's marketing claim of "open source."
12 agents in this collection
Best for Engineering teams building a bespoke, production-grade agent who want the largest integration ecosystem plus real stateful control.
The most widely adopted agent framework here; MIT-licensed core, free to self-host with no vendor lock-in.
Best for Developers who want to stand up a role-based, collaborating multi-agent system quickly in Python, with a stable MIT-licensed core and a hosted build and runtime for production.
MIT-licensed multi-agent orchestration framework, reached a stable 1.0 general-availability release in October 2025.
Best for Python or .NET developers and researchers exploring multi-agent orchestration who want a mature, free, self-hosted framework.
Microsoft's MIT-licensed multi-agent framework, now frozen in favor of its own successor below; free to self-host, pay only your own model provider.
Best for .NET or Python teams already invested in Azure or Microsoft Foundry who want one supported agent SDK, an optional managed hosting path, and a documented migration off AutoGen or Semantic Kernel.
Microsoft's MIT-licensed successor to AutoGen and Semantic Kernel, free to self-host across .NET, Python and Go, or an optional Azure-billed managed layer on top.
Best for Engineers building RAG-heavy or document-heavy agents who want an MIT-licensed OSS core with an optional managed parsing and extraction platform on top.
MIT-licensed core framework for building agents over your own data; the paid LlamaCloud parsing layer is optional.
Best for Engineering teams building production agents who need cyclical graphs, durable state and replayable runs
MIT-licensed agent runtime from the LangChain team, free to self-host; LangSmith and LangGraph Platform are optional hosted layers on top, not a gate on the library itself.
Best for Teams that want Google's own official, actively-developed agent SDK, especially polyglot teams (Python/Go/Java/TypeScript/Kotlin) or anyone building multi-agent systems that need to interoperate across vendors via the A2A protocol.
Apache-2.0 licensed across all four of its language SDKs (Python, Go, Java, TypeScript); Google's own official framework, free to self-host anywhere.
Best for Developers who want the fastest, lightest way to ship a straightforward agent (support triage, a tool-using assistant or a sandboxed coding agent) with tracing and guardrails included.
MIT-licensed and free to self-host, OpenAI's own lightweight SDK for building agentic workflows.
Best for Python teams building type-safe agents who also want realtime voice, image generation and embeddings in one SDK, with optional durable execution via Temporal, DBOS, Prefect or Restate.
MIT-licensed, type-safe Python framework from the team behind Pydantic; free to run anywhere.
Best for Developers and researchers who specifically need an agent whose memory, skills and identity genuinely persist and evolve across long-running sessions, rather than reset on every call.
Apache-2.0 licensed framework with persistent memory; self-host free or opt into managed Letta Cloud later.
Best for Python developers who want an Apache 2.0 SDK with memory, knowledge, learning, guardrails and a Control Plane they run inside their own cloud account, and who are comfortable running Docker or a cloud deploy template themselves.
Apache-2.0 licensed Python framework, formerly Phidata, with built-in memory, knowledge and multi-agent Teams; the paid AgentOS runtime is a separate BYOC product, not a gate on the framework itself.
Best for JS/TS developers who want an ownable, dependency-light agent toolkit
MIT-licensed TypeScript foundation for building agents, with the newest and least proven track record here; AgentsKit OS, a separate hosted layer, remains enterprise-only early access, but the open-source core is the only way most teams will run it.