Agent Frameworks
Frameworks and platforms for building AI agents.
Tool | Category | Segment | Framework / Platform | Plan | Monthly Price USD | Pricing Model | Free Tier / Trial | Included Runs / Usage | Primary Language / Runtime | Agent / Workflow Capabilities | Tool / MCP / API Support | State / Memory / Persistence | Observability / Evals | Deployment / Hosting | Team / Governance | Best Fit | Main Limits / Caveats |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
No tagline | Agent Frameworks | Graph-based agent framework | LangGraph | Open source | $0 software | Open-source framework; model and infra costs separate | ✓ | No LangGraph software usage meter when self-hosted; pay model provider, database, queue and hosting costs | Python and JavaScript/TypeScript | Stateful graphs, multi-step agents, multi-agent workflows, human-in-the-loop, persistence and streaming | Tool calling through LangChain/LangGraph nodes; MCP and deployed API patterns through LangSmith platform | Checkpointing and stores via local/DB-backed implementations; app owns persistence in OSS | Local tracing/evals require added tools such as LangSmith, Langfuse or custom telemetry | Self-host anywhere; deploy as your own service | No built-in hosted team governance in OSS | Teams needing explicit control over long-running stateful agent workflows | Framework is powerful but lower-level than role-based multi-agent frameworks; production ops are yours |
No tagline | Agent Frameworks | Managed agent platform | LangSmith / LangGraph Platform | Plus | $39/seat/mo | Seat plus usage for traces, deployments, Fleet, Engine and Sandboxes | No; paid plan | 10K base traces/month; 1 free dev-sized agent deployment; 500 Fleet runs/month; unlimited Fleet agents | Hosted platform for Python/JS LangGraph agents | Deploy and manage long-running agents; Fleet agents; Engine and Sandboxes access | 30+ deployment API endpoints, MCP server exposure and remote MCP tools per pricing page | LangSmith Cloud; hybrid/self-hosted options require Enterprise | Tracing, evals, monitoring, annotation queues, Engine issue diagnosis and sandbox usage | Cloud; production uptime billed per minute; deployment runs billed after included dev deployment | Unlimited seats; up to 3 workspaces; email support | Teams shipping LangGraph agents with hosted deployment and observability | Model usage billed separately; production uptime/runs/sandboxes add metered costs |
No tagline | Agent Frameworks | Managed agent platform | LangSmith / LangGraph Platform | Enterprise | Custom | Custom enterprise package | No public self-serve free tier beyond Developer | Custom seats, workspaces, Fleet, Engine and Sandbox packages | Hosted/hybrid/self-hosted LangSmith platform | Enterprise deployment, observability, evals, Fleet, Engine and Sandboxes | LangGraph APIs, MCP server exposure, remote MCP tools and custom packages | Cloud, hybrid or self-hosted; self-hosted data plane/VPC options | Enterprise tracing/evals/monitoring and support SLA | Cloud, hybrid or fully self-managed | Custom SSO, RBAC, SLA, trainings and architectural guidance | Enterprises standardizing on LangGraph with strict data/governance needs | Requires sales; costs depend on hosting model and usage |
No tagline | Agent Frameworks | Role-based multi-agent framework | CrewAI | Open source | $0 software | Open-source Python framework; model and infra costs separate | ✓ | No framework usage meter when run locally/self-hosted; you pay LLM/tool/hosting costs | Python | Agents, crews, flows, tasks, processes, memory, knowledge, guardrails and human-in-the-loop patterns | CrewAI tools, LangChain tools, custom tools and framework adapters | Memory and knowledge features available in framework; app owns storage/runtime | OpenTelemetry, tracing, LLM testing and guardrails via platform/docs depending setup | Local/self-hosted or own cloud | Open-source project governance; team controls depend on your infra | Builders who want role/task-oriented multi-agent orchestration without hand-authoring graph state machines | Abstractions can be less explicit than graph-first orchestration for complex deterministic flows |
No tagline | Agent Frameworks | Hosted agent workflow platform | CrewAI | Free | $0 | Hosted workflow execution plan | ✓ | 50 workflow executions/month; unlimited deployments listed | CrewAI cloud / AMP-style hosted platform | Visual editor, AI copilot, deployments, tracing, OpenTelemetry, LLM testing, guardrails and human input | GitHub integration; integration toolkit; Slack/Teams workflow chat requires Enterprise | Hosted CrewAI cloud; enterprise/private infra only on Enterprise | Tracing, performance metrics, token count, hallucination scores and deployment history | CrewAI cloud | Community support; enterprise adds RBAC/SSO | Testing CrewAI workflows and validating whether AMP fits a team | Hard 50 execution cap; no dedicated support or enterprise controls |
No tagline | Agent Frameworks | Hosted agent workflow platform | CrewAI | Enterprise | Custom | Enterprise hosted/private infrastructure plan | Request trial | Unlimited maximum executions; included executions sized to workflow; flexible overage | CrewAI cloud or customer/private infrastructure | Everything in Free plus enterprise deployment, scale, dedicated support/training and workflow chat integrations | GitHub integration, enterprise integrations, Slack/Teams support and private infrastructure options | CrewAI or customer infrastructure; dedicated VPC, NAT, FedRAMP High options shown | Tracing, OpenTelemetry, usage dashboard, LLM testing, guardrails and metrics | CrewAI cloud, private infrastructure or customer environment | SSO, RBAC, dedicated Slack/Teams support, onsite/training/development support | Large organizations operationalizing agentic workflows with support and private infrastructure | No public list price; execution sizing and overage are sales-led |
No tagline | Agent Frameworks | API-native agent SDK | OpenAI Agents SDK | SDK | $0 SDK; API/tool usage billed | Open-source SDK plus OpenAI API token/tool pricing | SDK is free; API requires paid usage unless account credits apply | No SDK run quota; usage billed through model tokens and tools such as web search, file search and containers | Python and TypeScript | Agents, specialist handoffs, guardrails, human review, streaming, traces, voice agents and sandbox agents | Function tools, hosted tools, MCP integrations and OpenAI platform tools | Application owns state/storage; SDK supports resumable state and run results | Built-in tracing for LLM generations, tool calls, handoffs and guardrails | Self-host orchestration; controlled sandbox capabilities available via API/SDK | Governance depends on API org/project controls and app implementation | Apps already using OpenAI models that want a native, typed agent loop with handoffs and tracing | Tightly aligned with OpenAI APIs; model/tool costs can dominate, and long-lived infrastructure is app-owned |
No tagline | Agent Frameworks | API-native agent runtime | OpenAI API Tools | Usage-based | Usage-based | Per-token and per-tool API pricing | No standing free tier listed on pricing page | Examples: web search $10/1K calls; containers from $0.03 per 20-min 1GB session; file search $0.10/GB/day after 1GB free plus $2.50/1K tool calls | OpenAI API runtime | Model calls plus hosted tools for agent workflows | Web search, hosted shell/code interpreter containers, file search and other platform tools | File storage/tool state billed separately; app controls broader persistence | Usage can be traced through Agents SDK and platform logs | OpenAI-hosted tools plus your application server | API account/project controls; enterprise controls depend on contract | Budgeting OpenAI-native agents beyond SDK-level software cost | Prices vary by selected model, context length and tools; verify before high-volume workloads |
No tagline | Agent Frameworks | Open-source agent SDK | Google ADK | Open source | $0 software | Open-source framework; model/cloud costs separate | ✓ | No ADK framework quota when run locally; pay Gemini/API/cloud/runtime costs separately | Python, TypeScript, Go, Java and Kotlin | Agents, graph workflows, tools, memory, callbacks, plugins and multi-agent patterns | Built-in tools, function tools, third-party tools, MCP tools and REST/API references | Memory and persistence patterns in ADK; deployment runtime can use Google Cloud services | Evaluation/debug/dev UI capabilities in ADK ecosystem | Local, self-hosted or Google Cloud/Vertex AI Agent Engine | Enterprise controls come from Google Cloud when deployed there | Teams wanting a Google-backed, multi-language agent SDK optimized for Gemini but not limited to local prototyping | Best fit if you are comfortable with Google/Gemini ecosystem; cloud deployment costs are separate |
No tagline | Agent Frameworks | Managed agent runtime | Vertex AI Agent Engine | Usage-based | Usage-based after free tier | Compute, memory, sessions and memory-bank usage | Monthly runtime free tier | Runtime free tier: first 180,000 vCPU-seconds and 360,000 GiB-seconds per month/project; then vCPU $0.0864/hr and RAM $0.009/GiB-hr | Google Cloud managed runtime for ADK and agent apps | Scale agents in production; runtime plus code execution, sessions and memory bank services | Google Cloud APIs, Vertex AI SDK, tool integrations and cloud IAM | Sessions billed per stored event; memory stored/retrieved billed separately after free retrieval allowance | Cloud monitoring/logging and Vertex AI tooling | Google Cloud managed runtime | IAM, regions, VPC/security controls depend on Google Cloud configuration | Deploying ADK/agent workloads to managed Google Cloud infra with a small runtime free tier | Code execution, sessions and memory bank started billing from Jan 28, 2026; model calls billed separately |
No tagline | Agent Frameworks | Microsoft agent SDK | Microsoft Agent Framework | Open source | $0 software | Open-source SDK; model/cloud costs separate | ✓ | No framework usage meter; pay Azure OpenAI/OpenAI/Azure AI/model and hosting costs separately | .NET and Python | Unified foundation for agents and graph-based workflows, extending ideas from Semantic Kernel and AutoGen | Tools, Agent Skills, providers, A2A protocol and Azure Functions/Durable Task integrations | Memory and persistence patterns; app/runtime owns storage | DevUI and integration with Microsoft/Azure observability patterns | Self-hosted app, Azure Functions, Durable Task or other .NET/Python hosting | Enterprise governance via Azure/Microsoft stack | Teams in Microsoft/.NET/Azure ecosystems building future-facing agent workflows | Newer unified framework; migration may be needed from AutoGen/Semantic Kernel projects |
No tagline | Agent Frameworks | Legacy/open multi-agent framework | Microsoft AutoGen | Open source | $0 software | Open-source framework; model/cloud costs separate | ✓ | No framework usage meter; pay model/tool/hosting costs | Python | Multi-agent conversation framework with AgentChat, teams, extensions and AutoGen Studio ecosystem | OpenAI/Azure/OpenAI-compatible model clients and extension APIs | Application/runtime owns persistence; newer Microsoft Agent Framework is successor direction | AutoGen Studio and custom logging/tracing depending setup | Self-hosted/local | Open-source community; enterprise governance through your infrastructure | Existing AutoGen projects and teams that need stable multi-agent conversations | Microsoft now positions Agent Framework as the unified successor; consider migration for new Microsoft-stack projects |
No tagline | Agent Frameworks | Microsoft LLM app framework | Semantic Kernel | Open source | $0 software | Open-source SDK; model/cloud costs separate | ✓ | No framework usage meter; pay model/tool/hosting costs | C#, Python and Java | Agents, plugins, planning/orchestration, memory patterns and function/tool composition | Plugins expose app functions to LLM workflows; connectors to OpenAI, Azure OpenAI and other services | Memory/connectors available; app owns persistence | Observability depends on Microsoft/Azure/OpenTelemetry setup | Self-hosted app or Azure/Microsoft app stack | Governance through app/Azure controls | Teams with existing .NET/Java/Python business logic to expose as agent tools | Microsoft Agent Framework is the forward-looking unified path for agents/workflows |
No tagline | Agent Frameworks | TypeScript agent framework | Mastra | Open source | $0 software | Apache-2.0 open-source framework; model/infra costs separate | ✓ | No OSS framework usage meter; self-host and pay provider/hosting/storage costs | TypeScript / JavaScript | Agents, workflows, tools, memory, instructions, durable multi-step processes and observational memory | MCP, tools, providers and deployable server/runtime concepts | Memory and storage abstractions; app owns infrastructure in OSS | Observability hooks; hosted platform adds event metering and retention | Host anywhere | Enterprise self-hosted features require custom license | TypeScript teams wanting a code-first agent framework with workflows and memory | Hosted platform features are separate from OSS; memory gateway/model costs are separate |
No tagline | Agent Frameworks | Hosted agent platform | Mastra Platform | Starter | $0/mo | Hosted platform with included events/CPU and overages | ✓ | 100K observability events then $10/100K; 24 CPU hours then $0.35/hr; 15-day retention; 1GB platform storage | Mastra hosted platform | Hosted agents, workflows, Studio, Server and Memory Gateway | Provider/model access via Memory Gateway; BYOK supported; MCP/tools via framework | 1GB data storage; 250MB retrieval storage then $20/GB | 100K logs/metrics/traces included; 15-day retention | Mastra hosted platform | Unlimited users, deployments and projects; no multiple-team features | Free hosted start for TypeScript agent projects with visible observability and CPU quotas | Persistent server costs $100/project; memory/model gateway usage has separate market-rate/markup charges |
No tagline | Agent Frameworks | Hosted agent platform | Mastra Platform | Teams | $250/mo | Hosted platform with included events/CPU and overages | No; paid plan | 1M observability events then $8/100K; 250 CPU hours then $0.25/hr; 6-month retention; 100GB platform storage | Mastra hosted platform | Multiple teams, SSO, SOC 2 docs, hosted agents/workflows and Server | MCP/tools/providers through Mastra framework and gateway | 100GB storage; 1GB retrieval storage then $20/GB | 1M logs/metrics/traces included; 6-month retention | Mastra hosted platform; on-prem/custom via Enterprise | Multiple teams, SSO and SOC 2 docs | Growing teams standardizing on Mastra with hosted observability/runtime | CPU/events/storage overages apply; persistent server and memory gateway costs are separate |
No tagline | Agent Frameworks | Document ingestion cloud for agents | LlamaIndex Cloud / LlamaParse | Starter | $50/mo | Credit-based document parsing/indexing plan | No; paid plan | 40K credits/month; pay-as-you-go up to 400K credits; 5 users | LlamaIndex Cloud service | Document ingestion/parse/extract/index services for RAG/agent workflows | Cloud API and LlamaIndex integrations | Hosted document/indexing service | Basic support | LlamaIndex Cloud | 5 users | Small teams processing documents for RAG agents | Not a full managed agent runtime; credits can be consumed quickly by heavy parsing |
No tagline | Agent Frameworks | Document ingestion cloud for agents | LlamaIndex Cloud / LlamaParse | Pro | $500/mo | Credit-based document parsing/indexing plan | No; paid plan | 400K credits/month; pay-as-you-go up to 4,000K credits; 10 users; Slack support | LlamaIndex Cloud service | Higher-volume document ingestion for production RAG/agent workflows | Cloud API and LlamaIndex integrations | Hosted service; enterprise adds SaaS/hybrid deployment | Slack support | LlamaIndex Cloud | 10 users | Teams running significant document ingestion behind agents | Full enterprise governance/5x rate limits require Enterprise |
No tagline | Agent Frameworks | AI gateway for Pydantic AI | Pydantic AI Gateway / Logfire | Gateway | Pricing through Logfire/account | Gateway/BYOK or managed-provider inference | Sign-up based; no static public price captured in docs | Usage depends on Logfire/Gateway provider billing, cost limits and chosen managed/BYOK providers | Pydantic AI with Logfire gateway | Single API key across multiple providers, cost limits and native provider formats | OpenAI, Anthropic, Google Vertex, Groq, AWS Bedrock and more via gateway | Gateway account manages API keys and spend limits; app owns application state | Logfire observability and enterprise features such as SSO/custom roles | Hosted Logfire gateway | Enterprise features inherit from Logfire | Teams wanting provider routing and spend controls around Pydantic AI agents | Docs describe gateway capabilities but do not list static plan prices on this page |
No tagline | Agent Frameworks | Minimal Python code-agent framework | Hugging Face smolagents | Open source | $0 software | Open-source Python library; model/hosting/sandbox costs separate | ✓ | No framework usage meter; pay for Hugging Face Inference Providers, API models, local compute or sandbox providers as used | Python | CodeAgent, ToolCallingAgent, multi-step agents, multi-agent orchestration examples, vision/web agents and CLI tools | Tools from MCP servers, LangChain, Hub Spaces and custom Python tools | Agent memory/persistence is app-owned; Hub/Spaces can host shared agents/tools | Logging/evals depend on surrounding stack | Local, Docker, Gradio Spaces, Modal, Blaxel, E2B or own infra | Governance depends on hosting/provider | Developers wanting a small, transparent Python agent library with code-first tool execution | Code execution needs sandboxing; production governance is not a full platform |
No tagline | Agent Frameworks | LLM program optimization framework | DSPy | Open source | $0 software | Open-source framework; model/infra costs separate | ✓ | No framework usage meter; pay model calls during runtime and optimization/evaluation | Python | Signatures, modules, ReAct/tool modules, composition, multimodal fields and optimizers for prompts/programs | Tools as Python functions; model/provider adapters through DSPy LM configuration | Program state is code/app-owned; no managed memory platform by default | Optimization/evaluation metrics are core framework concepts | Self-hosted Python app or notebook/service | Governance depends on app/runtime | Teams optimizing LLM pipelines and agents against metrics rather than hand-tuning prompts | Not a turnkey agent hosting platform; optimization can increase model-call cost during compile/eval |
No tagline | Agent Frameworks | Python agent runtime/framework | Agno | Free / OSS | $0 | Open-source framework and local AgentOS control plane | ✓ | No per-event fees for local/open-source usage; run AgentOS locally and pay model/infra costs | Python | Agents, teams, workflows, memory, knowledge, storage, guardrails, human-in-the-loop and AgentOS UI | 100+ integrations; can serve LangGraph, DSPy and Claude Agent SDK agents through AgentOS | Memory, knowledge and storage abstractions plus session history | Session monitoring, metrics and system evaluations in Free pricing copy | Local AgentOS or self-hosted services | Community support/forums on Free | Python teams wanting framework plus local operational UI for agents/teams/workflows | Live production control plane requires Pro; exact OSS license/limits should be checked in repo for redistribution |
No tagline | Agent Frameworks | Managed agent control plane | Agno AgentOS | Pro | $150/mo | Hosted/live AgentOS control plane subscription | No; paid plan | 1 live connection; 4 total seats included; unlimited monitoring, retention, knowledge, memories and chats; add seats/connections separately | AgentOS control plane for Python agent systems | Live management of agents, teams and workflows | AgentOS exposes/runtime-manages agent systems; external framework support noted in docs | Unlimited retention/knowledge/memories on pricing page | Unlimited monitoring | Hosted live AgentOS connection | $30/mo per extra seat; $95/mo per live connection | Managing production agent systems without per-event billing | Enterprise needed for self-hosted control plane, custom SSO/RBAC and mission-critical support |
No tagline | Agent Frameworks | Framework-agnostic TypeScript AI SDK | Vercel AI SDK | Open source | $0 SDK | Open-source SDK; gateway/hosting/model costs separate | ✓ | No SDK usage meter; pay model provider, AI Gateway, Vercel hosting/sandbox/workflow costs as used | TypeScript / JavaScript | Text/structured generation, tool calling, streaming, agent patterns and UI hooks | 16+ providers, AI Gateway, tools, structured outputs and framework adapters | Application owns persistence; Vercel ecosystem adds workflows/sandbox/hosting | DevTools and ecosystem observability; app-level tracing/evals depend on integrations | Any JS runtime; optimized for Vercel/Next.js | Governance depends on Vercel team/account and app code | Web app teams building streaming AI agents and chat/product UIs in TypeScript | More of an AI app/agent toolkit than a full multi-agent orchestration framework |
No tagline | Agent Frameworks | Open-source multi-agent research framework | CAMEL | Open source | $0 software | Open-source framework; model/infra costs separate | ✓ | No framework usage meter; pay model/tool/hosting costs | Python | Multi-agent society patterns, role-playing agents and agent scaling research/practice | Tools/connectors through CAMEL ecosystem and custom Python integration | Memory/state depends on app and CAMEL modules used | Research/eval tooling depends on setup | Self-hosted/local | Open-source community governance | Research and experimental multi-agent simulations or role-based collaboration | May require more engineering for production governance compared with managed platforms |
No tagline | Agent Frameworks | ModelScope multi-agent framework | AgentScope | Open source | $0 software | Open-source framework; model/infra costs separate | ✓ | No framework usage meter; pay model/tool/hosting costs | Python | LLM-empowered multi-agent applications, message passing, tools, workflows and examples | Custom tools, model services and ModelScope ecosystem integrations | Application-owned state/memory depending modules used | Logging/evaluation depends on framework modules and app setup | Self-hosted/local | Open-source community; enterprise controls are app-owned | Teams exploring open multi-agent systems in the ModelScope ecosystem | Less broadly adopted in Western SaaS stacks than LangGraph/CrewAI/OpenAI Agents SDK |