As of November 10, 2026, LangGraph is the best AI agent framework for developers building stateful, long-running agents: MIT licensed, Python and TypeScript, release 1.2.20 on November 5, with LangSmith tracing from $39 per seat a month. CrewAI is the fastest route to role-based multi-agent crews, the OpenAI Agents SDK the lightest, the Claude Agent SDK the pick for coding agents on Claude, and Mastra the TypeScript choice.
Eight code-first agent frameworks compared on license, language, release activity, model support and the paid hosting or tracing layer each one sells.
Most teams should start with LangGraph or CrewAI, and pick a vendor SDK only when they already know which model family they will live on. All eight frameworks here are free to install. What differs is how they model an agent (a graph, a crew of roles, a loop with handoffs), which languages they support, and which paid layer the company behind them sells for tracing and hosting.
The field moved a lot in 2026. Microsoft archived AutoGen and points new projects to Microsoft Agent Framework, which it also names as the successor to Semantic Kernel. The OpenAI Agents SDK now describes itself as provider-agnostic, with support for 100+ other LLMs. LlamaIndex says its company focus has shifted to LlamaParse, its document platform, while the open source framework stays available.
Toolradar data: our AI agents category lists 1,077 tools, and 56% of them have a free or freemium plan, yet every framework in this guide is free at the library level. Category trends sit on the AI agents statistics page.
How we ranked (methodology): eight code-first frameworks from that category, ordered by fit for agents that run in production (durable state, human review, tracing, model choice) and then by release activity on GitHub. License, language and release facts come from each project's own repository and docs; paid-layer prices from the vendors' pricing pages, as of November 10, 2026. Every total is our arithmetic.
Top Picks
Selected from the AI agents category for code-first frameworks a developer installs as a library, then ordered by fit for production agents (state, recovery, tracing, model choice) and by how actively each project ships releases. How we rate
Low-level orchestration framework for long-running, stateful agents, with durable execution, human-in-the-loop and memory.
Starting price
Free, MIT; LangSmith from $39/seat
Best for
Engineering teams building agents that run for a long time, branch, retry and pause for human review.
Python framework for multi-agent work: role-based Crews for autonomy and event-driven Flows for precise control.
Starting price
Free, MIT; platform free tier
Best for
Python teams that want a working multi-agent pipeline of specialist roles quickly.
Lightweight framework for multi-agent workflows with handoffs, guardrails, sessions, tracing, voice and sandbox agents.
Starting price
Free, MIT
Best for
Developers who want the smallest set of concepts between them and a working agent.
| Tool | Starting price | Rating | Best for |
|---|---|---|---|
| LangGraph | Free, MIT; LangSmith from $39/seat | n/a | Engineering teams building agents that run for a long time, branch, retry and pause for human review. |
| CrewAI | Free, MIT; platform free tier | n/a | Python teams that want a working multi-agent pipeline of specialist roles quickly. |
| OpenAI Agents SDK | Free, MIT | n/a | Developers who want the smallest set of concepts between them and a working agent. |
| Claude Agent SDK | Free SDK; Claude API usage | n/a | Teams building coding, file or operations agents on Claude. |
| Mastra | Free core; Platform Teams $250/mo | n/a | TypeScript and Node.js teams that want agents, workflows and memory in one framework. |
| Semantic Kernel | Free, MIT | 3.3 7 reviews | .NET and Java teams maintaining existing Semantic Kernel code in the Microsoft stack. |
| LlamaIndex | Free, MIT; LlamaParse credits extra | n/a | Teams whose agents mostly read and reason over a private document corpus. |
| AG2 (AutoGen) | Free, Apache 2.0 | n/a | Researchers and existing AutoGen users who want the conversational multi-agent model maintained. |
LangGraph
Low-level orchestration framework for long-running, stateful agents, with durable execution, human-in-the-loop and memory.
- Starting price
- Free, MIT; LangSmith from $39/seat
- Free plan
- Free to use
Best for: Engineering teams building agents that run for a long time, branch, retry and pause for human review.
- Pro: Durable execution resumes an agent exactly where it stopped after a failure.
- Pro: Python and JavaScript/TypeScript libraries, both MIT licensed.
- Pro: Frequent releases: 1.2.15 to 1.2.20 shipped between October 12 and November 5, 2026.
- Con: Low-level by design; LangChain points newcomers to Deep Agents, a higher-level package, for quick builds.
- Con: Tracing at team scale means paid LangSmith Plus seats, plus usage beyond 10,000 base traces a month.
- Starting price
- Free, MIT; platform free tier
- Free plan
- Yes
Best for: Python teams that want a working multi-agent pipeline of specialist roles quickly.
- Pro: Crews and Flows combine autonomous agents with deterministic, event-driven steps in one project.
- Pro: Platform free tier includes the visual editor, tracing, OpenTelemetry and guardrails.
- Pro: A workflow can be exported as an MCP server or a UI component.
- Con: Python only.
- Con: The free platform tier stops at 50 workflow executions a month, and the next step is Enterprise by quote.
CrewAI has two plans: a free Basic plan with 50 workflow executions per month, the visual editor, AI copilot and GitHub integration, and a custom-priced Enterprise plan.
OpenAI Agents SDK
Lightweight framework for multi-agent workflows with handoffs, guardrails, sessions, tracing, voice and sandbox agents.
- Starting price
- Free, MIT
- Free plan
- Free to use
Best for: Developers who want the smallest set of concepts between them and a working agent.
- Pro: Provider-agnostic: OpenAI's Responses and Chat Completions APIs plus 100+ other LLMs.
- Pro: Built-in tracing, guardrails, sessions and human-in-the-loop with no extra service.
- Pro: Python and JavaScript/TypeScript versions, both MIT.
- Con: Still pre-1.0 (v0.24.0 on October 15, 2026), so APIs can change between releases.
- Con: Multi-agent work runs through handoffs and agents-as-tools, with no graph of cycles like LangGraph's.
The OpenAI Agents SDK is entirely free and open-source, making it an incredibly generous offering.
Claude Agent SDK
Anthropic's SDK that puts Claude Code's agent loop, built-in tools, hooks and subagents inside your Python or TypeScript app.
- Starting price
- Free SDK; Claude API usage
- Free plan
- Free to use
Best for: Teams building coding, file or operations agents on Claude.
- Pro: Ships Claude Code's built-in tools: read, write and edit files, run commands and search the web.
- Pro: Hooks, permissions, sessions, subagents and MCP are part of the SDK.
- Pro: Very active: Python v0.2.180 released November 9, 2026.
- Con: Built around Claude models, so switching providers means another framework.
- Con: Governed by Anthropic's Commercial Terms rather than a plain open source license.
The Anthropic Agent SDK is offered as a free, open-source solution, which is incredibly generous and fair compared to any market alternative.
Mastra
TypeScript framework for agents and durable workflows with model routing to 40+ providers and memory built in.
- Starting price
- Free core; Platform Teams $250/mo
- Free plan
- Yes
Best for: TypeScript and Node.js teams that want agents, workflows and memory in one framework.
- Pro: Model routing to 40+ providers through one interface.
- Pro: Can author MCP servers that expose its agents and tools to other systems.
- Pro: Mastra Platform Starter is free with 100,000 observability events and unlimited users.
- Con: TypeScript only, so Python-first data teams need a second stack.
- Con: Code in ee/ folders needs a Mastra Enterprise License for production use.
Mastra's pricing is extremely generous, offering a full 'Open Source' tier completely free.
Semantic Kernel
Microsoft's agent and orchestration SDK for C#, Python and Java, now succeeded by Microsoft Agent Framework.
- 3.3 on Trustpilot (7 reviews)
- Starting price
- Free, MIT
- Free plan
- Free to use
- Rating
- 3.3 (7 reviews)
Best for: .NET and Java teams maintaining existing Semantic Kernel code in the Microsoft stack.
- Pro: The only pick with Java support (JDK 17+), alongside .NET 10 and Python 3.11 to 3.14.
- Pro: Still shipping: .NET 1.82.0 released November 7, 2026.
- Pro: Microsoft publishes a migration guide to Microsoft Agent Framework.
- Con: Microsoft calls Agent Framework its successor, so new features land there first.
- Con: Python support starts at 3.11 and .NET at 10.0, so older runtimes need an upgrade first.
LlamaIndex
Open source framework for retrieval-heavy agents over private documents, with Workflows and data connectors.
- 4.8 on G2 (2 reviews)
- Starting price
- Free, MIT; LlamaParse credits extra
- Free plan
- Yes
Best for: Teams whose agents mostly read and reason over a private document corpus.
- Pro: Workflows and retrieval primitives for agents that reason over private data.
- Pro: Works with LlamaParse, which includes 10,000 free credits a month for new users.
- Pro: MIT licensed, with integrations for LLM, embedding and vector store providers.
- Con: The company says its focus has shifted to LlamaParse, and the last framework release was v0.14.25 on September 21, 2026.
- Con: Parsing beyond the free credits costs $1.5 per 1,000 credits on pay-as-you-go.
LlamaIndex offers a very generous open-source tier, providing a full LLM framework and RAG capabilities for free.
AG2 (AutoGen)
Community-maintained continuation of AutoGen for agents that cooperate through conversation, while Microsoft's AutoGen repo is archived.
- Starting price
- Free, Apache 2.0
- Free plan
- Free to use
Best for: Researchers and existing AutoGen users who want the conversational multi-agent model maintained.
- Pro: AG2 v1.3.0 shipped November 4, 2026, maintained by volunteers from several organizations.
- Pro: AG2 Classic keeps the original autogen import, GroupChat and ConversableAgent working in a separate repo.
- Pro: Human-in-the-loop and multi-agent conversation patterns are core features.
- Con: Microsoft's AutoGen is archived with no bug or security fixes, and Microsoft advises against production use.
- Con: AG2 v1.0 moved to a new ag2 import, so older AutoGen code needs AG2 Classic or a rewrite.
Other AI Agents tools worth considering
More published tools from our AI Agents category, ordered by our category ranking. They are not part of the editorial picks above.
What is an AI agent framework?
An AI agent framework is a code library that runs the loop behind an agent: the model plans a step, calls a tool, reads the result and decides what to do next until the task is done. The framework handles state, tool calls, retries, memory and tracing, so your code describes what the agent should do rather than how to keep it running.
That is the difference from no-code agent builders, which trade control for speed. A framework lives in your repository, runs in your process and can be stepped through in a debugger, which matters once an agent touches customer data or runs unattended for hours.
Frameworks also differ from the Model Context Protocol (MCP). MCP standardizes how an agent reaches tools and data; the framework decides when to call them. Most frameworks in this guide can act as MCP clients, and several (Mastra, CrewAI, the Claude Agent SDK) can expose agents or tools over MCP too.
Which AI agent framework should a developer pick?
Pick by orchestration model and language first, then check the paid layer you will need in production. The table sets the eight side by side as of November 10, 2026.
| Framework | Orchestration model | Languages | Model support | Paid layer |
|---|---|---|---|---|
| LangGraph | Graph with cycles, durable execution | Python, JavaScript/TypeScript | Any model your nodes call | LangSmith tracing, evals and deployment |
| CrewAI | Crews of role-based agents plus event-driven Flows | Python | OpenAI by default; local models via Ollama and more | CrewAI platform: free tier, then Enterprise by quote |
| OpenAI Agents SDK | Agents with handoffs, guardrails and sessions | Python, JavaScript/TypeScript | OpenAI plus 100+ other LLMs | None; you pay model usage |
| Claude Agent SDK | Claude Code's agent loop with built-in tools | Python, TypeScript | Claude | None; you pay Claude API usage |
| Mastra | Agents plus typed, durable workflows | TypeScript | 40+ providers | Mastra Platform, free tier or $250 a month |
| Semantic Kernel | Agents with plugins, memory and planning | C#, Python, Java | Model-agnostic, incl. Ollama and ONNX | None |
| LlamaIndex | Retrieval-first agents and Workflows | Python | LLM, embedding and vector store integrations | LlamaParse credits for document parsing |
| AG2 (AutoGen line) | Conversations among agents | Python | Various LLMs | None |
What a paid layer costs, worked through. Take a team of three developers on LangSmith Plus whose agents produce 30,000 traces a month. Three Plus seats cost $117. Plus includes 10,000 base traces, and each extra trace is 0.005 LangChain Standard Units at $1 per unit, so the other 20,000 traces add $100. That is $217 a month, by our arithmetic, before any model API spend. The same team on Mastra Platform Starter pays nothing up to 100,000 observability events and 24 CPU hours, then $10 per extra 100,000 events.
Also worth knowing: Microsoft Agent Framework (MIT, Python and .NET, release python-1.25.0 on November 7, 2026) is the successor Microsoft names for both AutoGen and Semantic Kernel, and Google's Agent Development Kit (Apache 2.0, v2.14.0 on November 7, 2026) is another actively released open source option. Neither is ranked here yet.
Key Features to Look For
Orchestration model (Essential)
Graphs (LangGraph), role-based crews (CrewAI), handoff loops (OpenAI Agents SDK) or conversations (AG2) decide how much control you have over each step.
Durable state and recovery (Essential)
LangGraph resumes a long-running agent where it stopped after a failure; check that your choice persists state outside process memory.
Language support (Essential)
Python is universal; TypeScript teams look at Mastra, LangGraph.js, the OpenAI JS SDK or the Claude Agent SDK; C# and Java point to Semantic Kernel.
Model choice (Important)
Mastra routes to 40+ providers and the OpenAI Agents SDK to 100+ LLMs, while the Claude Agent SDK is built around Claude.
Human-in-the-loop (Important)
LangGraph, the OpenAI Agents SDK, CrewAI and AG2 document ways to pause an agent for review or approval.
Tracing and evals (Important)
Built-in tracing (OpenAI Agents SDK, CrewAI with OpenTelemetry) or a paid service (LangSmith, Mastra Platform) shows why an agent took each step.
MCP support (Nice to have)
Agents reach tools and data through MCP servers; Mastra can also author MCP servers, and CrewAI can export a workflow as one.
How to choose for your team
-
Long-running agents that must survive crashes and pause for human approval: LangGraph.
-
A pipeline of specialist roles (researcher, writer, reviewer) shipped quickly in Python: CrewAI.
-
A small agent with handoffs and guardrails, on OpenAI or another provider: the OpenAI Agents SDK.
-
A coding or file-system agent with built-in read, edit and shell tools: the Claude Agent SDK.
-
A TypeScript product team: Mastra, or LangGraph.js if you want LangGraph's model.
-
An existing .NET or Java codebase: Semantic Kernel today, with a plan to move to Microsoft Agent Framework.
Evaluation Checklist
Read the repository status first: an archived project (AutoGen) gets no security fixes.
Check the date of the latest release, not only the star count.
Confirm the license fits your product: the Claude Agent SDK is governed by Anthropic's Commercial Terms, and Mastra's ee/ folders need an enterprise license in production.
Build one real task end to end, including a failure and a retry, before you commit.
Price the tracing you will need at production volume, not at prototype volume.
List which model providers you might move to within a year and check the framework supports them.
Pricing Overview
Open source libraries
LangGraph, CrewAI, the OpenAI Agents SDK, Semantic Kernel, LlamaIndex and AG2 under MIT or Apache 2.0, plus Mastra's Apache 2.0 core.
Free
Free hosted tiers
LangSmith Developer (1 seat, 5,000 base traces a month), Mastra Platform Starter (100,000 events) and CrewAI Free (50 executions a month).
$0
Team tracing and hosting
LangSmith Plus with 10,000 base traces and one free serverless deployment, or Mastra Platform Teams with 1 million events and SSO.
Per seat, or flat monthly
Document parsing for retrieval agents
LlamaParse pay-as-you-go after 10,000 free credits, for agents that read PDFs and scans.
$1.5 per 1,000 credits
Enterprise
LangSmith Enterprise self-hosting, CrewAI Enterprise in your own VPC and Mastra Enterprise licensed features.
Contact sales
Pricing Comparison
| Framework | License | Latest release on GitHub | Paid layer and entry price |
|---|---|---|---|
MIT | 1.2.20, Nov 5, 2026 | LangSmith: Developer $0; Plus $39/seat/mo | |
MIT | 1.15.39, Nov 9, 2026 | Platform Free (50 executions/mo); Enterprise by quote | |
MIT | Python v0.24.0, Oct 15, 2026 | None; model usage billed by your provider | |
Anthropic Commercial Terms | Python v0.2.180, Nov 9, 2026 | None; Claude API usage | |
Mastra | Apache 2.0 core; ee/ under enterprise license | @mastra/core 1.82.0, Nov 4, 2026 | Platform Starter $0; Teams $250/mo |
MIT | .NET 1.82.0, Nov 7, 2026 | None; successor is Microsoft Agent Framework | |
MIT | v0.14.25, Sep 21, 2026 | LlamaParse: 10K free credits, then $1.5 per 1K | |
AG2 (AutoGen line) | Apache 2.0 | v1.3.0, Nov 4, 2026 | None |
Releases and licenses as of November 10, 2026 from LangGraph, CrewAI, OpenAI Agents SDK, Claude Agent SDK and its docs, Mastra, Semantic Kernel, LlamaIndex, AG2, AutoGen and Microsoft Agent Framework. Prices from LangSmith pricing, CrewAI pricing, Mastra pricing and LlamaParse pricing. Totals are our arithmetic.
Mistakes to Avoid
- ×
Starting a new project on Microsoft's AutoGen without noticing the repository is archived and gets no security fixes.
- ×
Choosing a framework by GitHub stars instead of the date of its latest release.
- ×
Pricing tracing at prototype volume: 10,000 base traces a month on LangSmith Plus disappear fast once agents run in production.
- ×
Assuming every SDK is MIT: the Claude Agent SDK follows Anthropic's Commercial Terms and part of Mastra uses an enterprise license.
- ×
Coupling business logic to one framework's classes, which makes a later move to a successor (such as Microsoft Agent Framework) a rewrite.
Expert Tips
- →
Keep tools behind MCP servers so a framework change does not mean rewriting every integration.
- →
For TypeScript, compare Mastra with LangGraph.js before committing; both are free and actively released.
- →
If you run CrewAI in production, wire its OpenTelemetry output into the tracing stack you already pay for.
- →
Weighing role-based crews against conversational agents? AutoGen vs CrewAI sets them side by side, and CrewAI alternatives covers the rest.
- →
Before production, read the AI agent observability guide and the AI agent testing guide; LangGraph alternatives helps if its graph model feels too low-level.
Red Flags to Watch For
-
A repository marked archived or in maintenance mode, or a README pointing new users to a successor.
-
No release in the last few months for a framework you plan to run in production.
-
A license that is not a standard open source license, with no note on commercial use.
-
Tracing or hosting priced per trace with no published rate.
-
A framework that cannot persist agent state outside process memory.
The Bottom Line
LangGraph is the best AI agent framework for most developers because it handles the hard production problems (durable state, recovery, human review) and ships core releases several times a month. CrewAI is the faster start for role-based multi-agent work in Python, and Mastra is the pick for TypeScript teams.
Choose the OpenAI Agents SDK for a light, provider-agnostic agent and the Claude Agent SDK for coding and file agents on Claude. Treat Semantic Kernel and AutoGen as migration questions now that Microsoft names Agent Framework as their successor. For no-code and business agent platforms, see the AI agent platforms guide; for open source agents you can run as-is, the open source AI agents guide.
Frequently Asked Questions
What are the best AI agent frameworks for developers in 2026?
As of November 10, 2026, LangGraph is the best for stateful, long-running agents (MIT, Python and TypeScript, release 1.2.20 on November 5). CrewAI is best for role-based multi-agent crews in Python, the OpenAI Agents SDK for a light provider-agnostic agent, the Claude Agent SDK for coding agents on Claude, and Mastra for TypeScript teams.
How much does it cost to run an AI agent framework?
The frameworks are free; you pay for model usage and, optionally, tracing or hosting. LangSmith Plus is $39 per seat a month with 10,000 base traces, so three developers at 30,000 traces a month pay $217, by our arithmetic. Mastra Platform Starter is free and Teams is $250 a month; CrewAI's platform has a free tier of 50 executions a month.
Is there a free AI agent framework?
Yes, all eight in this guide are free to use. LangGraph, CrewAI, the OpenAI Agents SDK, Semantic Kernel and LlamaIndex are MIT licensed, AG2 and Mastra's core are Apache 2.0, and the Claude Agent SDK is free to install under Anthropic's Commercial Terms. Free hosted tiers also exist: LangSmith Developer, Mastra Platform Starter and CrewAI Free.
LangGraph vs CrewAI: which should I pick?
Pick CrewAI when the work splits into clear roles that hand results to each other and you want a working Python pipeline quickly. Pick LangGraph when the agent needs cycles, durable state that survives a crash, or a pause for human approval at any step. Both are MIT licensed and released in November 2026.
Is AutoGen still maintained?
No. Microsoft archived AutoGen, provides no bug or security fixes and advises against production use; it points new users to Microsoft Agent Framework. AG2, a community-maintained continuation, released v1.3.0 on November 4, 2026, and keeps the original autogen import in AG2 Classic.
Does the OpenAI Agents SDK only work with OpenAI models?
No. Its own README describes it as provider-agnostic, supporting OpenAI's Responses and Chat Completions APIs plus 100+ other LLMs. It is MIT licensed, available for Python and JavaScript/TypeScript, and was at v0.24.0 on October 15, 2026.
What is the best AI agent framework for TypeScript?
Mastra is built for TypeScript, with agents, durable workflows, memory and routing to 40+ model providers, under an Apache 2.0 core. LangGraph.js, the OpenAI Agents SDK for JavaScript and the Claude Agent SDK for TypeScript are the other options, depending on whether you want a graph model or a vendor SDK.
Cite this page: Toolradar, "Best AI Agent Frameworks in 2026", updated November 2026, https://toolradar.com/guides/best-ai-agent-frameworks
Spotted an outdated price or a wrong fact? Suggest a correction: we check it against the vendor's own page. Work for one of these vendors? Claim your listing to update it yourself.
Related Guides
From the team behind Toolradar
Editorial content for AI startups
We turn AI product expertise into content that ranks, gets cited by LLMs, and reaches 950K+ tech buyers.
See how we workReady to Choose?
Compare features, read reviews, and find the right tool.
Some offers on this page may be paid placements or contain affiliate links.