Geekflare

Best Open Source AI Agent Frameworks in 2026

I looked at 53 open source frameworks for building AI agents and ranked the ones I'd pick. I checked which models it works with, whether a run can pause for approval and resume after a crash, and which parts need the vendor's paid cloud. Then I built the same research agent in seven of them, LangGraph, CrewAI, Pydantic AI, the OpenAI Agents SDK, Mastra, Microsoft Agent Framework and Google ADK.

Data refreshedMethodologyHow we evaluate
Director of Geekflare

Rity Anand is a director of Geekflare. She has extensive experience managing business technology and a special interest in emerging technologies.

What I found

  • All seven frameworks I tested answered the same question correctly in all three runs, with the same model and tools. For a simple agent, the model matters more than the framework.
  • Mastra, the OpenAI Agents SDK and Microsoft Agent Framework gave the quickest answers, at about 20 seconds each, while Pydantic AI took 38 seconds.
  • Microsoft Agent Framework and CrewAI were the heaviest installs, at 78 and 57 seconds and close to 400 MB each, and CrewAI cost about twice as much per answer as Mastra.
  • Using a model through OpenRouter needed a fix in three of them: Pydantic AI's default output limit, a missing package in Google ADK, and Microsoft Agent Framework's default client, which uses an OpenAI-only API.

Tested on DigitalOcean Droplet with Ubuntu 24.04 server with 2 vCPU and 4 GB of RAM, using Docker. “Ready in” runs from an empty server to the first working screen or result, image downloads included. The research task gives every framework the same question, the same model, Claude Haiku 5.5 through OpenRouter, and the same web search and page-reading tools, built on Geekflare's APIs. Each ran three times and the time shown is the median. These are libraries that run inside your own app, so there is no idle memory to report.

SoftwareVersionReady inResearch task, median of 3What happened
LangGraph1.2.1645 s24.8 s, 133 MB peakPassed all three runs with the prebuilt agent in a few lines. Without LangSmith there's no trace viewer, so I read the steps from the messages it returns.
CrewAI1.15.281 min 23 s29.6 s, 391 MB peakThe most memory of the seven at 390 MB during install and run, and about twice the cost per answer of the cheapest frameworks.
Pydantic AI2.57.054 s37.7 s, 132 MB peakThe quickest install, but by default it asks for 65,536 output tokens per call, which an account with little credit rejects. With a cap it passed all three runs.
OpenAI Agents SDK0.24.041 s20.6 s, 120 MB peakWorked with a non-OpenAI model through its chat completions class, but its tracing uploads to OpenAI by default, so I switched it off.
Mastra1.78.056 s20.3 s, 349 MB peakTied for the fastest median run and among the cheapest per answer. Its model router took an OpenRouter model name directly, with no provider package.
Microsoft Agent Framework1.22.01 min 38 s20.3 s, 362 MB peakThe slowest install here at 78 seconds, and its default OpenAI client uses an API that OpenRouter doesn't support. With the chat completions client it was among the fastest.
Google ADK2.12.01 min 17 s22.9 s, 349 MB peakWorked with a non-Gemini model through LiteLLM, but only after I installed the mcp package, which that route needs and the install doesn't pull in.

10 Open Source AI Agent Frameworks compared

Licence, languages, deployment options and the vendor's hosted plan side by side. Sort by latest release to see how often each one ships, since most of them release every week or two.

LicenceTypeDockerSSOManaged cloud
1
LangGraph
Python 3.10+ / TypeScript
MIT
Permissive
Agent orchestration framework43kOct 2026(Actively maintained)NoPaid only$39/seat/month
2
CrewAI
Python 3.10–3.13
MIT
Permissive
Multi-agent framework60kOct 2026(Actively maintained)NoPaid onlyFree for 50 runs a month
3
Pydantic AI
Python 3.11+
MIT
Permissive
Agent framework21kOct 2026(Actively maintained)NoNoNone
4
OpenAI Agents SDK
Python 3.10+ / TypeScript
MIT
Permissive
Agent SDK30kOct 2026(Actively maintained)NoNoNone
5
Mastra
TypeScript / Node.js 22.13+
Apache-2.0 and Mastra Enterprise License
Open core
TypeScript agent framework29kOct 2026(Actively maintained)NoFree editionFree Starter, Teams $250/month
6
Microsoft Agent Framework
Python 3.10+ / .NET
MIT
Permissive
Agent and workflow framework14kOct 2026(Actively maintained)NoNoBilled by container compute
7
Google ADK
Python 3.11+ / TypeScript / Java / Go
Apache-2.0
Permissive
Agent development kit22kOct 2026(Actively maintained)NoNo$0.085 per vCPU-hour
8
Agno
Python 3.9+ / FastAPI
Apache-2.0
Permissive
Agent framework and runtime43kOct 2026(Actively maintained)YesPaid only$150/month for a live deployment
9
LlamaIndex
Python 3.10+
MIT
Permissive
Agent and RAG framework53kSep 2026(Actively maintained)NoNo$50/month, free plan with 10K credits
10
Strands Agents
Python 3.10+ / TypeScript
Apache-2.0
Permissive
Agent SDK8.9kOct 2026(Actively maintained)YesNo$0.0895 per vCPU-hour
  • Actively maintained
  • Slowing: last release 1 to 2 years ago
  • Stale: no release in over 2 years

What each one can do

The features that decide whether an agent can go into production, checked against each project's docs. Paid means the feature is only in the vendor's paid plan or cloud.

SoftwareAny model providerMulti-agentMCP toolsDurable stateHuman approvalTracingVisual studio
LangGraphYesYesSubgraphs, supervisors and handoffsYesThrough the langchain-mcp-adapters packageYesCheckpointers for SQLite and PostgresYesPartlyLangSmith cloud, free for 5,000 traces a monthPartlyLangSmith Studio, a hosted page that connects to your server
CrewAIYesYesYesYesCheckpoints to JSON files or SQLiteYesPartlyBuilt-in traces go to CrewAI AMP; others via OpenTelemetryPartlyCrew Studio is in the hosted CrewAI AMP, with a free plan
Pydantic AIYesYesDelegation through tools, handoffs or graphsYesYesVia Temporal, DBOS, Prefect or RestateYesYesOpenTelemetry; Logfire is optional and hostedPartlyA web chat UI for trying agents, no builder
OpenAI Agents SDKYesOther providers through LiteLLM or any-llmYesHandoffs and agents as toolsYesPartlyPaused runs save as RunState; crash recovery via Temporal or DBOSYesYesOpenAI's Traces dashboard by default, or your own processorNo
MastraYesYesYesYesWorkflow snapshots in LibSQL, Postgres or other storageYesYesShown in Studio, with OpenTelemetry exportersYesLocal Studio; team roles need the enterprise licence
Microsoft Agent FrameworkYesYesSequential, concurrent, handoff and group chat patternsYesYesWorkflow checkpointsYesYesOpenTelemetry to any backendYesDevUI, a local web UI, in beta
Google ADKYesGemini first; others through LiteLLM, Claude and Ollama connectorsYesYesYesResumable runs in Python and Kotlin; sessions in a databaseYesTool confirmation, still marked experimentalYesOpenTelemetry to any backendYesadk web dev UI, with a visual builder in Python
AgnoYesYesYesPartlyRuns and sessions saved; crashed runs restart from the startYesYesOpenTelemetry, with traces stored in your databasePartlyHosted AgentOS UI, free for local, $150/month for live
LlamaIndexYesYesYesThrough llama-index-tools-mcpYesWorkflow checkpoints, SQLite store or DBOSYesYesInstrumentation module with an OpenTelemetry exporterPartlyDebugger UI in the workflow server, no builder
Strands AgentsYesBedrock by default; Anthropic, OpenAI, Gemini, Ollama and moreYesYesPartlySession managers save state to files or S3YesYesOpenTelemetry, on by defaultNo
Ask Sparkian

Which agent framework fits your project?

Describe what the agent should do, the language your team writes and where it will run, and AI will shortlist the best fits from these picks.

Or try:

Agent frameworks I'd recommend

Ranked on what a team needs to run agents it can trust: control over each step, recovery from failures, human approval, tracing it can keep in-house, and how much of that is free.

  1. #1Best for stateful agent workflowsTested by Geekflare
    Full reviewWebsite GitHub

    LangGraph

    Graph-based framework for building long-running, stateful agents in Python and JavaScript

    The most control over how an agent moves between steps, with checkpoints and interrupts in the free library. Tracing, Studio and the production server lean on LangSmith.

    Tested Oct 19, 2026, version 1.2.16
    Ready in
    45 s
    Research task, median of 3
    24.8 s, 133 MB peak

    Passed all three runs with the prebuilt agent in a few lines. Without LangSmith there's no trace viewer, so I read the steps from the messages it returns.

    Commands, notes and screenshots

    Best for

    Teams building long-running agents that must pause, resume and be audited step by step

    Not for

    Teams that want a quick agent in a few lines or no tie to a vendor's platform for running it

    • Durable execution with checkpoints in SQLite or Postgres
    • Interrupts that pause a run for a person to approve or edit state
    • Works with any model provider, with or without LangChain
    • Python and JavaScript versions
    • Low-level, so simple agents take more code than in other frameworks
    • Tracing and Studio depend on LangSmith, a hosted service
    • The production Agent Server is not open source and needs a licence key to self-host
    Licence
    MIT (permissive)
    Stack
    Python 3.10+ / TypeScript
    Databases
    PostgreSQL, SQLite
    GitHub
    43k stars · 290 contributors
    Latest release
    1.2.16 · Oct 16, 2026
    Managed cloud
    LangSmith, $39/seat/month
    SSO
    Paid only
    Audit logs
    Not documented
  2. #2Best for teams of agents with little codeTested by Geekflare
    Full reviewWebsite GitHub

    CrewAI

    Python framework for teams of role-based AI agents and event-driven flows

    Describe each agent by role and goal and CrewAI handles the hand-offs, with flows and checkpoints for the fixed parts. Its tracing and visual editor are in the hosted CrewAI AMP.

    Tested Oct 19, 2026, version 1.15.28
    Ready in
    1 min 23 s
    Research task, median of 3
    29.6 s, 391 MB peak

    The most memory of the seven at 390 MB during install and run, and about twice the cost per answer of the cheapest frameworks.

    Commands, notes and screenshots

    Best for

    Teams that want several specialised agents working on one task with little code

    Not for

    Teams that need fine control over every step, or tracing without signing up to a vendor's platform

    • Agents defined by role, goal and backstory, so a working crew takes little code
    • Checkpointing lets a crew or flow resume after a failure
    • MCP servers can be added to an agent in one line
    • Standalone, with no LangChain dependency
    • Built-in tracing needs a CrewAI AMP account, though other tools work through OpenTelemetry
    • Sends anonymous telemetry unless you turn it off
    • Supports Python 3.10 to 3.13 only
    Licence
    MIT (permissive)
    Stack
    Python 3.10–3.13
    Databases
    SQLite
    GitHub
    60k stars · 349 contributors
    Latest release
    1.15.28 · Oct 14, 2026
    Managed cloud
    CrewAI AMP, Free for 50 runs a month
    SSO
    Paid only
    Audit logs
    Not documented
  3. #3Best for typed Python agentsTested by Geekflare
    Full reviewWebsite GitHub

    Pydantic AI

    Typed Python agent framework from the Pydantic team, with validated outputs and any model

    Agents that behave like ordinary typed Python, with validated outputs, almost any model, and durable runs through Temporal, DBOS or Prefect.

    Tested Oct 18, 2026, version 2.57.0
    Ready in
    54 s
    Research task, median of 3
    37.7 s, 132 MB peak

    The quickest install, but by default it asks for 65,536 output tokens per call, which an account with little credit rejects. With a cap it passed all three runs.

    Commands, notes and screenshots

    Best for

    Python teams that want typed, testable agents and structured outputs they can trust

    Not for

    Teams that want a visual builder or a ready-made multi-agent team abstraction

    • Outputs are validated against Pydantic models
    • Works with nearly every model provider and local servers
    • First-party durable execution with Temporal, DBOS, Prefect and Restate
    • OpenTelemetry tracing that works with any backend
    • Multi-agent setups are built by hand from delegation, handoffs or graphs
    • Only a simple web chat UI, no visual builder or debugger
    • Ships often, so pin versions and read the changelog when you upgrade
    Licence
    MIT (permissive)
    Stack
    Python 3.11+
    GitHub
    21k stars · 632 contributors
    Latest release
    v2.57.0 · Oct 16, 2026
    SSO
    No
    Audit logs
    No
  4. #4Simplest SDK to start withTested by Geekflare
    Full reviewWebsite GitHub

    OpenAI Agents SDK

    Lightweight SDK from OpenAI for agents with tools, handoffs, guardrails and tracing

    A small API of agents, tools, handoffs and guardrails, with tracing that works from the first run. Traces go to OpenAI unless you change that, and crash recovery needs an outside engine.

    Tested Oct 19, 2026, version 0.24.0
    Ready in
    41 s
    Research task, median of 3
    20.6 s, 120 MB peak

    Worked with a non-OpenAI model through its chat completions class, but its tracing uploads to OpenAI by default, so I switched it off.

    Commands, notes and screenshots

    Best for

    Developers who want a small, readable agent loop with handoffs and tracing from day one

    Not for

    Teams that need built-in crash recovery, a visual builder or tracing that stays off OpenAI's platform by default

    • A small API that is quick to learn
    • Tracing of every model call, tool call and handoff with no setup
    • Tool approvals that pause a run and resume it later from saved state
    • Works with other providers through LiteLLM or any-llm
    • Traces go to OpenAI by default, even for other providers' models
    • Crash recovery needs an outside engine such as Temporal or DBOS
    • No visual studio or server of its own
    Licence
    MIT (permissive)
    Stack
    Python 3.10+ / TypeScript
    Databases
    SQLite, Redis
    GitHub
    30k stars · 407 contributors
    Latest release
    v0.24.0 · Oct 15, 2026
    SSO
    No
    Audit logs
    No
  5. #5Best for TypeScript teamsTested by Geekflare
    Full reviewWebsite GitHub

    Mastra

    TypeScript framework for agents and workflows, with a local Studio, memory and tracing

    Agents, workflows that suspend and resume, memory and a local Studio in one TypeScript package. A few team features need its enterprise licence.

    Tested Oct 19, 2026, version 1.78.0
    Ready in
    56 s
    Research task, median of 3
    20.3 s, 349 MB peak

    Tied for the fastest median run and among the cheapest per answer. Its model router took an OpenRouter model name directly, with no provider package.

    Commands, notes and screenshots

    Best for

    TypeScript and Next.js teams building agents into their own product

    Not for

    Python teams, or teams that want every feature under an open source licence

    • Local Studio to chat with agents, run workflows and read traces
    • Workflows suspend for approval and resume from stored state
    • One interface for many model providers
    • Runs inside Next.js or Node apps, or as its own server
    • Some team features are under a source-available enterprise licence
    • Needs a recent Node.js, 22.13 or later
    • Many packages release separately, so upgrades touch several versions
    Licence
    Apache-2.0 and Mastra Enterprise License (open core)
    Stack
    TypeScript / Node.js 22.13+
    Databases
    LibSQL, PostgreSQL
    GitHub
    29k stars · 698 contributors
    Latest release
    1.78.0 · Oct 16, 2026
    Managed cloud
    Mastra platform, Free Starter, Teams $250/month
    SSO
    Free edition
    Audit logs
    Paid only
  6. #6Best for .NET and Azure teamsTested by Geekflare
    Full reviewWebsite GitHub

    Microsoft Agent Framework

    Microsoft's framework for agents and multi-agent workflows in Python and .NET, the successor to AutoGen and Semantic Kernel

    The successor to AutoGen and Semantic Kernel, in Python and .NET, with workflows, checkpoints, approvals and OpenTelemetry all free. Many samples start from Azure.

    Tested Oct 19, 2026, version 1.22.0
    Ready in
    1 min 38 s
    Research task, median of 3
    20.3 s, 362 MB peak

    The slowest install here at 78 seconds, and its default OpenAI client uses an API that OpenRouter doesn't support. With the chat completions client it was among the fastest.

    Commands, notes and screenshots

    Best for

    .NET and Python teams on Azure, and anyone moving off AutoGen or Semantic Kernel

    Not for

    TypeScript teams, or teams that want a hosted visual builder

    • Python and .NET with consistent APIs
    • Graph workflows with checkpointing, human-in-the-loop and time travel
    • OpenTelemetry tracing built in, with no paid service needed
    • Migration guides from AutoGen and Semantic Kernel
    • Most samples and quickstarts start from Azure and Foundry
    • DevUI is still a beta package
    • No TypeScript version, and Go is in preview
    Licence
    MIT (permissive)
    Stack
    Python 3.10+ / .NET
    GitHub
    14k stars · 298 contributors
    Latest release
    1.22.0 · Oct 15, 2026
    Managed cloud
    Microsoft Foundry hosted agents, Billed by container compute
    SSO
    No
    Audit logs
    No
  7. #7Best for Google Cloud and multi-language teamsTested by Geekflare
    Full reviewWebsite GitHub

    Google ADK

    Google's Agent Development Kit for building, testing and deploying agents in Python, TypeScript, Java and Go

    The same toolkit in Python, TypeScript, Java and Go, with a local dev UI, resumable runs and evaluations. It's tuned for Gemini.

    Tested Oct 19, 2026, version 2.12.0
    Ready in
    1 min 17 s
    Research task, median of 3
    22.9 s, 349 MB peak

    Worked with a non-Gemini model through LiteLLM, but only after I installed the mcp package, which that route needs and the install doesn't pull in.

    Commands, notes and screenshots

    Best for

    Teams on Google Cloud, and teams that need the same framework in Python, Java, Go or TypeScript

    Not for

    Teams that want a framework with no lean towards one model vendor's cloud

    • The same framework in Python, TypeScript, Java and Go
    • Local web UI with event history, state inspection and a visual builder
    • Tool confirmation and resumable runs built in, though confirmation is still experimental
    • Evaluation tools for agent trajectories
    • Tuned for Gemini, with other models through LiteLLM or connectors
    • The web UI is for development only, and the visual builder is Python only
    • Versions differ by language, and features land in Python first
    Licence
    Apache-2.0 (permissive)
    Stack
    Python 3.11+ / TypeScript / Java / Go
    Databases
    PostgreSQL, MySQL, SQLite
    GitHub
    22k stars · 594 contributors
    Latest release
    v2.12.0 · Oct 13, 2026
    Managed cloud
    Agent Runtime on Google Cloud, $0.085 per vCPU-hour
    SSO
    No
    Audit logs
    No
  8. #8Best for running agents as a service
    Website GitHub

    Agno

    Python framework and runtime for running agents, teams and workflows as a service in your own cloud

    One runtime gives you an API, storage, approvals and schedules on your own servers. Its UI is a hosted control plane that's paid for live deployments.

    Best for

    Python teams that want agents served as an API with storage, approvals and scheduling already built

    Not for

    Teams that need a self-hosted UI for production without a subscription

    • AgentOS serves agents as a REST API with streaming
    • Sessions, memory and traces stay in your own database
    • Approvals and cron scheduling built in
    • JWT-based RBAC and multi-tenant isolation in the free runtime
    • The UI is hosted by Agno and paid for live deployments
    • Sends a telemetry event per run unless you turn it off
    • A crashed background run restarts from the beginning rather than the last step
    Licence
    Apache-2.0 (permissive)
    Stack
    Python 3.9+ / FastAPI
    Databases
    PostgreSQL, SQLite, Redis
    GitHub
    43k stars · 609 contributors
    Latest release
    v3.1.5 · Oct 19, 2026
    Managed cloud
    AgentOS control plane, $150/month for a live deployment
    SSO
    Paid only
    Audit logs
    Free edition
  9. #9Best for agents over your documents
    Website GitHub

    LlamaIndex

    Python framework for agents and event-driven workflows over your documents and data

    The strongest retrieval and document tooling here, plus event-driven workflows that checkpoint, wait for people and run behind a server with a debugger.

    Best for

    Teams building agents that search, read and extract from their own documents

    Not for

    Teams that want a small agent loop without the retrieval stack

    • The most complete retrieval and document tooling of the frameworks here
    • Workflows that checkpoint and resume after a restart
    • A workflow server with a built-in debugger UI
    • OpenTelemetry tracing through an official package
    • Many separate integration packages to install and keep in step
    • Much of the newer agent tooling is built around documents and LlamaParse
    • The agent workflow packages release from a separate repo, so versions are spread out
    Licence
    MIT (permissive)
    Stack
    Python 3.10+
    Databases
    SQLite
    GitHub
    53k stars · 2k contributors
    Latest release
    v0.14.25 · Sep 21, 2026
    Managed cloud
    LlamaParse platform, $50/month, free plan with 10K credits
    SSO
    No
    Audit logs
    No
  10. #10Best for AWS teams
    Website GitHub

    Strands Agents

    AWS-backed open source SDK for model-driven agents in Python and TypeScript, with any model and no hosted control plane

    A lean, model-driven agent loop with OpenTelemetry tracing on by default and multi-agent patterns built in. It defaults to Bedrock but works with other providers.

    Best for

    Teams on AWS, and teams that want a lean, model-driven agent loop they can trace end to end

    Not for

    Teams that want a visual studio or a graph-first workflow engine

    • Works with Bedrock, Anthropic, OpenAI, Gemini, Ollama and others
    • OpenTelemetry tracing on by default, with no paid service
    • Swarm, graph and workflow patterns for several agents
    • Turn limits, token budgets and cancellation built in
    • Defaults to Amazon Bedrock, and the deployment guides are mostly AWS
    • No visual studio or dev UI
    • Persistence is session-based, not a step-by-step workflow checkpoint
    Licence
    Apache-2.0 (permissive)
    Stack
    Python 3.10+ / TypeScript
    GitHub
    8.9k stars · 334 contributors
    Latest release
    1.61.1 · Oct 15, 2026
    Managed cloud
    Amazon Bedrock AgentCore Runtime, $0.0895 per vCPU-hour
    SSO
    No
    Audit logs
    No

More open source agent frameworks

Frameworks for other languages, research projects and smaller libraries that fit a narrower need.

SoftwareLicenceTypeGitHub starsLatest release
LangChain
Python framework for building LLM agents that work with any model provider, built on LangGraph
MIT
Permissive
Agent framework148k(Actively maintained)Oct 2026
Haystack
Python framework for agents and RAG pipelines with explicit control over retrieval, routing and tools
Apache-2.0
Permissive
LLM orchestration framework27k(Actively maintained)Oct 2026
smolagents
Small Python library from Hugging Face for agents that write their actions as Python code
Apache-2.0
Permissive
Agent framework30k(Actively maintained)May 2026
AG2
Python framework for agents that cooperate, from the community fork of AutoGen 0.2
Apache-2.0
Permissive
Multi-agent framework5k(Actively maintained)Oct 2026
Semantic Kernel
Microsoft SDK for building agents and multi-agent workflows in C#, Python and Java
MIT
Permissive
Agent SDK29k(Actively maintained)Oct 2026
DSPy
Stanford framework for writing LM programs in Python and tuning their prompts automatically, with a ReAct agent module
MIT
Permissive
LLM programming framework39k(Actively maintained)Oct 2026
CAMEL
Multi-agent framework from the CAMEL-AI research community for agent societies, workforces and data generation
Apache-2.0
Permissive
Multi-agent framework18k(Actively maintained)Mar 2026
AI SDK
TypeScript toolkit from Vercel for building AI apps and tool-using agents with any model provider
Apache-2.0
Permissive
TypeScript AI toolkit27k(Actively maintained)Oct 2026
VoltAgent
TypeScript framework for agents, multi-agent teams and workflows, with an optional hosted console
MIT
Permissive
Agent framework11k(Actively maintained)Sep 2026
Genkit
Google's open source framework for AI apps and agents in TypeScript, Go, Python and Dart
Apache-2.0
Permissive
AI app and agent framework6.5k(Actively maintained)Sep 2026
Cloudflare Agents SDK
TypeScript SDK for stateful agents that run on Cloudflare Workers and Durable Objects
MIT
Permissive
Agent SDK for Cloudflare Workers5.8k(Actively maintained)Oct 2026
BeeAI Framework
Framework for agents and multi-agent systems in Python and TypeScript, hosted by LF AI & Data
Apache-2.0
Permissive
Multi-agent framework3.4k(Actively maintained)Sep 2026
Atomic Agents
Lightweight Python agent framework built on Instructor and Pydantic, with typed inputs and outputs
MIT
Permissive
Agent framework6.3k(Actively maintained)Oct 2026
Langroid
Python multi-agent framework where agents work through tasks and pass messages to each other
MIT
Permissive
Multi-agent framework4.1k(Actively maintained)Oct 2026
Agency Swarm
Multi-agent framework built on the OpenAI Agents SDK, with agents that talk along defined communication flows
MIT
Permissive
Multi-agent framework4.6k(Actively maintained)Aug 2026
Upsonic
Python framework for autonomous agents and task-based agent teams, with built-in safety policies
MIT
Permissive
Agent framework8k(Actively maintained)May 2026
Swarms
Python framework for multi-agent workflows, from sequential and concurrent runs to agent graphs
Apache-2.0
Permissive
Multi-agent framework7.3k(Actively maintained)Oct 2026
PocketFlow
Minimal LLM framework in about 100 lines of Python that models agents and workflows as a graph of nodes
MIT
Permissive
Minimal LLM framework11k(Slowing: last release 1 to 2 years ago)Jul 2025
Spring AI
Spring framework for AI features and agent patterns in Java apps, with MCP support
Apache-2.0
Permissive
Java AI framework9.6k(Actively maintained)Aug 2026
LangChain4j
Java library for LLM apps and agents, with workflows, a supervisor agent and MCP tools
Apache-2.0
Permissive
Java AI library13k(Actively maintained)Oct 2026
Koog
JetBrains framework for building AI agents in Kotlin or Java, with graph workflows and persistence
Apache-2.0
Permissive
Agent framework4.6k(Actively maintained)Sep 2026
Embabel
JVM agent framework from the creator of Spring that plans actions toward goals, built on Spring AI
Apache-2.0
Permissive
Agent framework with planning4.5k(Actively maintained)Oct 2026
Rig
Rust library for LLM apps and agents with one interface over many model providers
MIT
Permissive
Rust AI library8.9k(Actively maintained)Oct 2026
  • Actively maintained
  • Slowing: last release 1 to 2 years ago
  • Stale: no release in over 2 years

Not included, and why

Well-known names that are visual builders rather than frameworks, projects that stopped or changed direction, and SDKs that depend on a closed runtime.

  • Dify, Langflow and n8n: Visual builders for AI workflows rather than code-first frameworks. Dify's licence adds limits on multi-tenant use and on removing its logo, and n8n uses its own Sustainable Use License.
  • Flowise: The team stopped development in July 2026 and archived the repository in August 2026.
  • AutoGPT: Now a low-code agent platform, and the platform code is under the PolyForm Shield licence. Only the classic agent is MIT.
  • AutoGen: Microsoft archived it and no longer fixes bugs or security issues. Microsoft Agent Framework replaces it, and AG2 continues the community fork.
  • Letta: Its original agent server was retired. The project now centres on Letta Code, a personal agent app that uses Letta Cloud by default.
  • Claude Agent SDK: The wrapper is MIT, but it bundles the Claude Code CLI under Anthropic's commercial terms and works only with Claude models.
  • MetaGPT: No release since March 2025, and the project's site now points to the team's hosted app builder.
  • SuperAGI: No commits since January 2025.
  • OpenAI Swarm: An experimental project that OpenAI says is replaced by the Agents SDK.
  • TaskWeaver: Archived by Microsoft.
  • Julep: Being rewritten as Julep 3, with only release candidates so far, and its hosted service has moved to a different product.
  • elizaOS: Now an agent operating system with its own app and cloud, and the 2.x line is still in beta.
  • Rasa: Rasa Open Source is in maintenance mode, and its newer CALM approach is in the paid Rasa Pro.
  • LiveKit Agents: A framework for realtime voice agents, a narrower job than the frameworks here.
  • txtai: Built around semantic search and an embeddings database rather than agents.
  • Inngest AgentKit: Under 1,000 GitHub stars, and no release since November 2025.
  • OWL: A task automation project built on CAMEL, which is listed instead.
  • Browser Use: An AI agent that operates a browser rather than a framework for building agents. See open source Web Scraping Tools

How to choose

Start from your team's language

Most frameworks here are Python first. TypeScript teams should look at Mastra, the Vercel AI SDK or VoltAgent, .NET teams at Microsoft Agent Framework, and Java teams at Spring AI, LangChain4j, Koog or Embabel. Google ADK covers Python, TypeScript, Java and Go.

Decide how much control you need

LangGraph, Microsoft Agent Framework and the workflow engines in ADK, Mastra and LlamaIndex make you lay out the steps, which is more work but predictable. CrewAI, the OpenAI Agents SDK, Strands and Agno let the model plan more, which is faster to build and harder to test.

Plan for long runs and approvals

If an agent waits hours for a person or runs for a long time, pick one that saves its state and resumes, such as LangGraph, CrewAI, Microsoft Agent Framework, Mastra or LlamaIndex. Pydantic AI and the OpenAI Agents SDK get the same through Temporal or DBOS, which is one more system to run.

Check where the traces go

Tracing is where the paid clouds come in. LangGraph points to LangSmith, CrewAI to CrewAI AMP and the OpenAI Agents SDK to OpenAI's dashboard by default. Microsoft Agent Framework, Google ADK, Strands, Pydantic AI and Mastra emit OpenTelemetry, so traces can go to a backend you already run.

How we evaluate

No vendor influences these pages. A vendor can pay for a Featured placement, labelled Sponsored, which never changes a rank or a review.

What qualifies
Code-first frameworks and SDKs for building AI agents, with an OSI open source licence or a clearly labelled source-available one, a release or commit in the last 12 months, and 1,000+ GitHub stars. Visual and no-code agent builders, voice-only frameworks, single-purpose agents such as coding or browser agents, and SDKs that only run on a proprietary runtime are not considered.
Where we look
GitHub, the pages that rank for each category, so both established and newer projects are considered.
How we rank
Fit for business and technical teams: maintenance activity, self-hosting effort, licence terms, and which features the paid editions lock. Payment never changes a ranking.
Where the data comes from
Stars, releases, contributors and licences come from GitHub. Editions, pricing and requirements come from each software's own docs.
Hands-on testing
We install several of the picks in each category on a cloud Ubuntu 24.04 server (2 vCPU, 4 GB RAM) with Docker, and record install time and idle memory.
Questions

Frequently asked questions

For agents that need to pause, resume and be checked step by step, LangGraph. For several agents working together with little code, CrewAI. For typed Python, Pydantic AI, and for TypeScript, Mastra. Teams on Azure or Google Cloud will find Microsoft Agent Framework and Google ADK fit best.

LangGraph is lower level. You draw the agent as a graph, and in return you get fine control, checkpoints and interrupts at any step. CrewAI is higher level. You describe agents by role and goal and it runs them as a team, which is quicker to build but harder to steer. Both are MIT, and both send you to their own hosted platform for built-in tracing.

Yes. The CrewAI framework is MIT and free to self-host, including crews, flows, checkpointing and MCP tools. CrewAI AMP, the hosted control plane with tracing, Crew Studio and managed deployment, is a separate product with a free plan of 50 runs a month and custom enterprise pricing. The framework also sends anonymous telemetry unless you set OTEL_SDK_DISABLED=true.

All ten picks can use tools from MCP servers. In LangGraph it comes through the langchain-mcp-adapters package and in LlamaIndex through llama-index-tools-mcp, while CrewAI, the OpenAI Agents SDK, Pydantic AI, Mastra, Google ADK, Microsoft Agent Framework, Agno and Strands support it directly. Mastra can also publish your own tools as an MCP server.

No. Every pick runs as a library inside your own service. What the paid platforms add is mostly hosted tracing, a visual studio and managed deployment. The one to watch is LangGraph's Agent Server, which is source-available and needs a LangSmith licence key to self-host in production, so many teams serve their graphs with their own API instead.