Skip to content
Short answer: In the Claude Agent SDK vs LangGraph choice, pick the Claude Agent SDK if you want an agent that reads files, runs commands and edits code on Claude models with little setup. Pick LangGraph if you want to draw the flow yourself as a graph of durable, stateful steps on any model. The OpenAI Agents SDK and Microsoft Agent Framework sit in between, and all four are free and MIT-licensed or close to it.
These four tools solve the same problem in different ways. Our LangGraph page, Claude Agent SDK page, OpenAI Agents SDK page and Microsoft Agent Framework page cover each on its own. This post sets them side by side so you can see which one fits your project.
LangGraph and the agent SDKs at a glance
| LangGraph | Claude Agent SDK | OpenAI Agents SDK | Microsoft Agent Framework | |
|---|---|---|---|---|
| Made by | LangChain | Anthropic | OpenAI | Microsoft |
| Core idea | Graph of nodes with durable state | Claude Code’s agent loop as a library | Few building blocks: agents, handoffs, guardrails | Agents plus graph workflows |
| Languages | Python, JavaScript/TypeScript | Python, TypeScript | Python, JavaScript/TypeScript | .NET, Python, Go (preview) |
| License | MIT | Anthropic Commercial Terms | MIT | MIT |
| Models | Any, through integrations | Claude models | OpenAI by default, others through adapters | Foundry, Anthropic, Azure OpenAI, OpenAI, Ollama and more |
| Install (Python) | pip install -U langgraph | pip install claude-agent-sdk | pip install openai-agents | pip install agent-framework |
| Cost | Free, model usage extra | Free SDK, Claude API usage extra | Free, model usage extra | Free, model usage extra |
What is LangGraph?
LangGraph is a low-level orchestration framework and runtime for stateful agents. You model an agent as a graph. Nodes do the work and edges decide what runs next. You can mix hand-written logic with model-driven decisions in one flow.
Its strengths are the ones long-running agents need. Durable execution lets an agent survive a failure and resume where it stopped. You can inspect and change state while it runs, and it supports short-term and long-term memory. It does not require LangChain. If you are weighing those two, read LangChain vs LangGraph.
What are the agent SDKs?
The three SDKs come from model and cloud vendors. Each ships a ready-made agent loop, so you write less orchestration code than in LangGraph. They differ in what they assume about your stack.
Claude Agent SDK
Anthropic describes it as the same tools, agent loop and context management that power Claude Code, programmable in Python and TypeScript. Built-in tools read, write and edit files, run commands and search the web. Hooks, subagents, sessions, permissions and MCP support come with it. If MCP is new to you, our guide to what MCP servers are explains it.
It runs Claude models. Besides an Anthropic API key, the quickstart lists Amazon Bedrock, Google Cloud’s Agent Platform and Microsoft Foundry as ways to authenticate. The SDK is governed by Anthropic’s Commercial Terms, and Anthropic says third-party products must use API key authentication rather than offering claude.ai login. Both packages bundle a Claude Code binary. The terminal tool built on the same loop is Claude Code.
OpenAI Agents SDK
OpenAI calls it a lightweight toolkit with very few abstractions. It has five core pieces: agents, handoffs (and agents as tools), guardrails, sessions and tracing. Guardrails run input checks in parallel with the agent so a bad request fails fast. It also has built-in MCP integration. It is the production-ready successor to the experimental Swarm project.
It uses the Responses API by default for OpenAI models. Other providers work through adapters such as LiteLLM. A JavaScript/TypeScript version exists under the MIT license and needs Node.js 22 or later.
Microsoft Agent Framework
Microsoft Learn calls it the direct successor to Semantic Kernel and AutoGen, built by the same teams. It combines AutoGen’s simple agent abstractions with Semantic Kernel’s enterprise features such as session state, type safety, middleware and telemetry. It adds graph-based workflows for explicit multi-agent paths.
It supports .NET and Python, with a Go version in public preview that lacks declarative agents, RAG and functional workflows. Providers include Microsoft Foundry, Anthropic, Azure OpenAI, OpenAI and Ollama. Migration guides cover Semantic Kernel and AutoGen.
How do they differ on control?
This is the real split. LangGraph hands you the graph. You define the state, the nodes and the edges. The framework gives you persistence, memory and human checkpoints. It takes longer to learn and gives you the most say over what happens at each step.
The OpenAI and Claude SDKs start from the agent loop. You give the agent instructions and tools, and it plans its own steps. That gets you a working agent faster, with less visibility into the exact path. The Claude SDK leans toward coding and file work. The OpenAI SDK leans toward handoffs between small agents.
Microsoft Agent Framework offers both. Its docs say to use an agent for open-ended tasks and a workflow when the process has well-defined steps, and they add that if you can write a function for the task, you should do that instead of using an AI agent.
Which model providers can you use?
LangGraph is the most neutral. It works with model providers through its integrations and has no home vendor, so you can switch models without changing the orchestration.
Microsoft Agent Framework is also broad, listing Foundry, Anthropic, Azure OpenAI, OpenAI and Ollama. The OpenAI Agents SDK works with other providers through adapters, but its defaults and best support target OpenAI models. The Claude Agent SDK is the most tied to one vendor: it runs Claude models, though you can reach them through several clouds.
If you expect to change model vendors, LangGraph or Microsoft Agent Framework carry the least lock-in.
Which languages do they support?
LangGraph, the OpenAI Agents SDK and the Claude Agent SDK all cover Python and JavaScript or TypeScript. Microsoft Agent Framework is the one to look at for C# and .NET, where it is generally available. It also lists Go, but only in preview.
If your team writes .NET services, Microsoft Agent Framework is the natural fit. For a Python or TypeScript team the language will not decide it.
What does each one cost?
All four frameworks are free to install. You pay for the models you call, and for hosting if you want it managed.
- LangGraph: free. LangChain’s hosted deployment runs through LangSmith, and Klyvoo’s LangGraph page notes it needs the Plus plan or higher. You can run LangGraph on your own servers without it.
- Claude Agent SDK: free SDK. You pay Claude API rates for usage, billed to your API key.
- OpenAI Agents SDK: free. There is no official hosting, so you deploy it yourself.
- Microsoft Agent Framework: free. Optional Microsoft Foundry hosting is usage-based and priced through Azure.
Where each one wins
- LangGraph: fine control, durable execution that resumes after failures, state you can inspect and change and no model lock-in.
- Claude Agent SDK: a coding and file agent out of the box, with built-in tools, hooks and subagents.
- OpenAI Agents SDK: the fewest concepts to learn, plus handoffs and parallel guardrails.
- Microsoft Agent Framework: .NET support, enterprise features and a migration path from Semantic Kernel or AutoGen.
What do none of them do?
None of them designs a good agent for you. You still write the tools, the prompts and the checks. You also have to test what the agent does. None removes model costs or guarantees correct answers. All of them change quickly, so pin your versions and read release notes before you upgrade. For a wider view, see our best AI agent frameworks page, or read how multi-agent systems work.
Which should you choose?
- You want a coding agent on Claude models: Claude Agent SDK.
- You want the simplest start with handoffs between agents: OpenAI Agents SDK.
- You write .NET or already use Semantic Kernel or AutoGen: Microsoft Agent Framework.
- You need long-running agents that resume after failures: LangGraph.
- You want to switch model vendors later: LangGraph.
- You are new to agents: read how to build an AI agent first.
FAQ
Is the Claude Agent SDK better than LangGraph?
Neither is better for everyone. The Claude Agent SDK gives you a working coding agent fast but runs Claude models. LangGraph gives you more control over state and flow and works with any model.
Is the OpenAI Agents SDK free?
Yes. The SDK is open source under the MIT license. You pay for the model API usage your agents generate, and OpenAI offers no official hosting for it.
Can the OpenAI Agents SDK use models other than OpenAI’s?
Yes. It lists non-OpenAI providers and adapters such as LiteLLM. Its defaults and best support still target OpenAI models.
Does Microsoft Agent Framework replace Semantic Kernel?
Microsoft Learn calls Agent Framework the direct successor to both Semantic Kernel and AutoGen, and publishes migration guides for each.
Does LangGraph need LangChain?
No. LangChain’s docs say you do not need LangChain to use LangGraph, though the two integrate.
Which one supports the most languages?
Microsoft Agent Framework lists .NET, Python and Go, though Go is in preview. LangGraph and the OpenAI and Claude SDKs cover Python and JavaScript or TypeScript.
Keep reading
Details here come from the Claude Agent SDK docs, the OpenAI Agents SDK docs and the LangGraph docs, checked in October 2026. To compare more frameworks, browse the Klyvoo AI agent frameworks page or the MCP servers directory.