Open Source · MIT License · Updated daily via GitHub Actions
A curated index of open-source agent frameworks with live GitHub stats. Stars, forks, language, license, and topics fetched daily. Filter by language or topic, sort by any column.
| Framework |
Stars |
Forks |
Language |
License |
Topics |
|---|---|---|---|---|---|
An AI agent framework is a library or platform for building software that uses an LLM to decide what to do next, then acts on that decision — calling tools, reading and writing state, and looping until a task is done. In practice that means three things: a way to define tools the model can call, a way to hold state across steps (the conversation, intermediate results, task memory), and a runtime loop that keeps calling the model until the task finishes or fails.
The term covers a wide range of designs. Some frameworks (LangChain, CrewAI) give you high-level abstractions — agents, roles, crews — and hide the loop underneath. Others (LangGraph) expose the control flow directly as a graph or state machine so you can route between steps, tools, and multiple agents explicitly. Others still (Mastra, Microsoft Agent Framework) are runtime-specific SDKs tied to a particular language ecosystem. The table above indexes all of them side by side with live GitHub signal — stars, forks, language, license — so you can compare by what actually matters for a production decision rather than by marketing copy.
Every framework above is tagged by design pattern, not just popularity. Use the Topic filter to isolate any of these — or read on for what each one actually means.
Manages control flow between LLM calls, tools, and steps — chains, graphs, or state machines that decide what runs next.
Coordinates several agents with distinct roles, prompts, and tools, communicating toward a shared goal.
Lightweight patterns for loosely-coupled agents that hand off tasks to each other without a central orchestrator.
Focused on retrieval-augmented generation — indexing, retrieving, and grounding agent output in external documents or data.
Drag-and-drop or visual-canvas tools for composing agent workflows without hand-writing the orchestration code.
Aimed at long-running, goal-directed agents that plan and execute multi-step tasks with minimal human intervention.
Specialized for agents whose primary tool is writing, running, and debugging code.
Published or maintained by Microsoft, including the Microsoft Agent Framework and its predecessors (AutoGen, Semantic Kernel).
Adds persistent or long-term memory to agents. See the Memory Frameworks index for a deeper, dedicated comparison.
Experimental or academic frameworks exploring new agent architectures, not yet aimed at production use.
These three terms get used interchangeably, but they describe different layers of the stack.
A library for building one or more LLM-driven agents: prompt templates, tool calling, and an execution loop. Examples: LangChain, CrewAI, Mastra.
The control-flow layer that decides which step, tool, or agent runs next — usually a graph, DAG, or state machine. LangGraph is the clearest example; the term is often used as a synonym for "agent framework" in casual usage.
The full production wrapper around a model or agent framework — world model, control state, verification, recovery, and observability. A harness can compile down to any of the frameworks above. Read the full breakdown →
Six things worth checking before committing to one, roughly in the order they tend to eliminate options.
Does it match your stack — Python, TypeScript, C#/.NET, Go? Switching later means a rewrite, not a config change.
Graph-based (LangGraph), role/crew-based (CrewAI), or a plain code-first agent loop? This decides how explicit your control flow will be.
MIT and Apache-2.0 are unrestricted for commercial use. GPL/AGPL and source-available licenses carry real obligations — check the License column above before adopting.
Built-in short/long-term memory, or does it expect you to bring your own (Mem0, Zep, Letta)? See the memory frameworks index.
Native tracing and eval hooks, or bolt-on? Debugging a multi-step agent without traces is largely guesswork.
Stars, forks, and issue activity in the table above are a faster proxy for production-readiness than a project's own marketing.
A library or platform for building software where an LLM decides what to do next and the framework executes that decision — calling tools, managing state across steps, and looping until the task completes. LangChain, CrewAI, LangGraph, and Mastra are all examples, though they differ in how much control-flow detail they expose versus abstract away.
In practice the terms overlap heavily. "Orchestration framework" more precisely refers to the control-flow layer — the graph or state machine deciding which step or agent runs next — while "agent framework" is the broader umbrella that also includes prompt templating, tool-calling conventions, and memory. Most tools people call agent frameworks include an orchestration layer internally.
A framework gives you the building blocks — agents, tools, an execution loop. A harness is the full production wrapper around those blocks: a world model, control state, verification before acting, recovery from failures, and observability across the whole run. A harness typically compiles to one or more agent frameworks rather than replacing them. See harness engineering for the full distinction.
Both terms are used for it in practice. LangGraph is a low-level graph/state-machine layer for controlling multi-step and multi-agent execution, which is why it's often described specifically as an orchestration framework — and it's frequently used underneath higher-level agent abstractions rather than on its own.
A single-agent framework runs one LLM loop with tools and memory. A multi-agent framework coordinates several agents, each with its own role, prompt, and tools, communicating toward a shared goal — useful when a task naturally splits into specialized sub-tasks (research, coding, review) rather than one generalist agent handling everything.
Depends entirely on the license, not on being "open source." MIT and Apache-2.0 — the majority of frameworks in this index — are unrestricted for commercial use. Copyleft licenses like GPL or AGPL carry obligations that can affect proprietary products. Always check the License column above rather than assuming.
Python has the largest share of the ecosystem, including LangChain, CrewAI, and LangGraph's primary SDK. TypeScript/JavaScript options include Mastra and LangGraph.js. C#/.NET is covered primarily by the Microsoft Agent Framework. Filter by Language in the table above to see current options per ecosystem.
Stars, forks, and topics are refreshed daily via GitHub Actions — the freshness indicator in the nav bar shows the age of the current data pull.