# AI agent frameworks

> An agent is a model in a loop with tools. A framework hands you that loop, plus the parts you'd otherwise rebuild: memory, handoffs, retries, human approval and tracing. LangGraph, CrewAI, Pydantic AI and Google ADK start in Python. Mastra is TypeScript. The OpenAI Agents SDK ships both. phpClaw brings the loop into PHP apps. guilds.run sits a level up: a self-hosted app for standing teams of agents you can inspect. All eight are open source and free to run. You pay for models, and for a hosted platform if you add one.

8 tools, in the order they appear in the Digidle Tools directory. This is not a ranking.

- Page: https://digidle.com/best/ai-agent-frameworks/
- All best-of lists: https://digidle.com/best/

## Tools

- [guilds.run](https://digidle.com/tools/guilds-run/): Open-source (AGPL-3.0) multi-agent teams you can inspect: debug every prompt, tool call, memory change and agent interaction, control permissions and pick the model per agent. Self-host free on your own machine with Docker Compose; hosted plan from $19/month coming soon. Official site: https://www.guilds.run. Alternatives: https://digidle.com/alternatives/guilds-run/.
- [phpClaw](https://digidle.com/tools/phpclaw/): Open-source (MIT) AI agent framework for PHP 8.1+: a ReAct loop with tools (allowlisted shell, HTTP with SSRF checks, sandboxed file access, database queries, code search), memory and guard chains that screen messages before the model sees them. Works with Anthropic, OpenAI, Groq, Gemini, Mistral, DeepSeek or Ollama; eight adapters for Laravel, Symfony, Drupal, Magento, WordPress/WooCommerce, Joomla, PrestaShop and OpenCart, and an MCP server package that exposes your tools to Claude Code or Cursor. Official site: https://phpclaw.ai. Alternatives: https://digidle.com/alternatives/phpclaw/.
- [LangGraph](https://digidle.com/tools/langgraph/): LangChain's low-level orchestration framework for long-running, stateful agents: durable execution that resumes where a run failed, human-in-the-loop checks on agent state, and short- and long-term memory. Python 3.10+ or LangGraph.js, with or without LangChain. Open source (MIT); deploy it yourself or on LangSmith Deployment. Official site: https://www.langchain.com/langgraph. Alternatives: https://digidle.com/alternatives/langgraph/.
- [CrewAI](https://digidle.com/tools/crewai/): Open-source (MIT) Python framework for multi-agent automation: Crews of role-based agents that work one task together, and Flows for event-driven control that mix single LLM calls with Crews. Standalone, not built on LangChain. CrewAI AMP adds managed deployment, tracing and governance; free plan with 50 workflow executions a month. Official site: https://crewai.com. Alternatives: https://digidle.com/alternatives/crewai/.
- [OpenAI Agents SDK](https://digidle.com/tools/openai-agents-sdk/): OpenAI's open-source (MIT) SDK for multi-agent workflows in Python or TypeScript: agents with tools, handoffs, guardrails, sessions and human-in-the-loop, built-in tracing, sandbox agents for long tasks, and voice agents. Works with the OpenAI APIs and 100+ other LLMs. Free; you pay for model usage. Official site: https://github.com/openai/openai-agents-python. Alternatives: https://digidle.com/alternatives/openai-agents-sdk/.
- [Mastra](https://digidle.com/tools/mastra/): TypeScript framework for AI agents and apps: agents with tools, graph-based workflows that suspend for human approval and resume, model routing across 40+ providers, built-in evals and observability, and Mastra Studio to test agents locally. Apache-2.0 outside its ee/ enterprise directories; Mastra Platform free to start, Teams $250/month. Official site: https://mastra.ai. Alternatives: https://digidle.com/alternatives/mastra/.
- [Pydantic AI](https://digidle.com/tools/pydantic-ai/): Typed Python agent framework from the Pydantic team: any model a string swap away with fallback, validated structured outputs, MCP, evals and durable execution on Temporal, DBOS, Prefect and more. One agent runs as a CLI, a web chat or a realtime voice agent. Open source (MIT) and free; Logfire tracing is optional. Official site: https://pydantic.dev/pydantic-ai. Alternatives: https://digidle.com/alternatives/pydantic-ai/.
- [Google ADK](https://digidle.com/tools/google-adk/): Google's open-source (Apache-2.0) Agent Development Kit: code-first agents and graph workflows, multi-agent hierarchies, MCP and OpenAPI tools, tool confirmation, a local dev UI and built-in evals. Optimized for Gemini but model-agnostic. Python, Java, Go, TypeScript and Kotlin; deploy with Docker, Cloud Run or Vertex AI Agent Engine. Official site: https://adk.dev. Alternatives: https://digidle.com/alternatives/google-adk/.

## At a glance

| Tool | What it is | Runs on | Open source | Price |
| --- | --- | --- | --- | --- |
| [guilds.run](https://digidle.com/tools/guilds-run/) | Standing teams of agents with roles, routines and shared tables. Every prompt, tool call and memory change is kept for inspection. | Your machine, as one Docker Compose stack, with your own model keys. Hosted version not open yet. | Yes, AGPL-3.0. | Self-hosted free. Hosted from $19/mo when it opens; model usage billed separately. |
| [phpClaw](https://digidle.com/tools/phpclaw/) | An agent engine for PHP: a ReAct loop with tools, memory, skills and eight guards that screen each message before the model sees it. | PHP 8.1+ on your own server. Adapters for Laravel, Symfony, WordPress, Drupal, Joomla, Magento, PrestaShop and OpenCart. | Yes, MIT. | Free. You pay your model provider. |
| [LangGraph](https://digidle.com/tools/langgraph/) | Low-level orchestration for long-running, stateful agents: durable execution, human-in-the-loop and memory. Works without LangChain. | Python 3.10+, or LangGraph.js. Host it yourself or on LangSmith Deployment. | Yes, MIT. | Free. LangSmith Developer $0 for 1 seat. Plus $39/seat/mo with 1 free small serverless deployment, then usage-based. |
| [CrewAI](https://digidle.com/tools/crewai/) | Multi-agent automation: Crews of role-based agents that share a task, and Flows for event-driven control. Standalone, not built on LangChain. | Python 3.10 to 3.13. CrewAI AMP runs it on CrewAI's cloud, your VPC or your own infrastructure. | Yes, MIT. | Framework free. AMP Basic free with 50 workflow executions a month. Enterprise custom. |
| [OpenAI Agents SDK](https://digidle.com/tools/openai-agents-sdk/) | OpenAI's lightweight multi-agent SDK: agents with tools, handoffs, guardrails, sessions and tracing, plus sandbox and voice agents. | Python 3.10+ or TypeScript. The OpenAI APIs or 100+ other LLMs. | Yes, MIT. | Free. You pay for the models you call. |
| [Mastra](https://digidle.com/tools/mastra/) | A TypeScript framework for agents and graph-based workflows, with model routing across 40+ providers, human-in-the-loop, evals and observability. | Node.js 22.13+. Inside React, Next.js or Node apps, as a standalone server, or on Mastra Platform. | Apache-2.0, except code in ee/ directories, which is under Mastra's enterprise license. | Framework free. Platform Starter $0/mo: 100K observability events, 24 CPU hours. Teams $250/mo. Enterprise custom. |
| [Pydantic AI](https://digidle.com/tools/pydantic-ai/) | A typed Python agent framework: swap models with a string, get validated outputs, plus MCP, evals and durable execution on Temporal, DBOS, Prefect and more. | Python 3.11+. OpenAI, Anthropic, Google, Bedrock, Groq, Mistral, Ollama and dozens more. | Yes, MIT. | Free. Optional Logfire tracing: Personal free up to 10M records a month, Team $49/mo. |
| [Google ADK](https://digidle.com/tools/google-adk/) | Google's code-first agent kit: agents and graph workflows, multi-agent hierarchies, tool confirmation, a dev UI and built-in evals. Built for Gemini, works with other models. | Python 3.11+, plus ADK for Java, Go, TypeScript and Kotlin. Deploy with Docker, Cloud Run or Vertex AI Agent Engine. | Yes, Apache-2.0. | Free. Google Cloud hosting and Gemini usage are billed by Google; no ADK-specific price. |

Checked 2026-10-31 on vendor pages: [LangGraph on GitHub](https://github.com/langchain-ai/langgraph), [LangSmith pricing](https://www.langchain.com/pricing), [CrewAI on GitHub](https://github.com/crewAIInc/crewAI), [CrewAI pricing](https://www.crewai.com/pricing), [OpenAI Agents SDK (Python)](https://github.com/openai/openai-agents-python), [OpenAI Agents SDK (TypeScript)](https://github.com/openai/openai-agents-js), [Mastra on GitHub](https://github.com/mastra-ai/mastra), [Mastra license](https://github.com/mastra-ai/mastra/blob/main/LICENSE.md), [Mastra pricing](https://mastra.ai/pricing), [Pydantic AI on GitHub](https://github.com/pydantic/pydantic-ai), [Logfire pricing](https://pydantic.dev/pricing), [Google ADK on GitHub](https://github.com/google/adk-python), [Google ADK docs](https://adk.dev), [phpClaw](https://phpclaw.ai), [phpClaw on GitHub](https://github.com/phpclaw-php/phpclaw-monorepo), [guilds.run pricing](https://www.guilds.run/pricing), [guilds.run on GitHub](https://github.com/CigaleAI/guilds).

## Pick it if

- [LangGraph](https://digidle.com/tools/langgraph/): You want explicit control over every step, and runs that survive a crash and pick up where they stopped.
- [CrewAI](https://digidle.com/tools/crewai/): You think in roles: a researcher, a writer and a reviewer, each with a job, working one task together.
- [OpenAI Agents SDK](https://digidle.com/tools/openai-agents-sdk/): You want the fewest moving parts: agents, handoffs and guardrails, with tracing built in, in Python or TypeScript.
- [Pydantic AI](https://digidle.com/tools/pydantic-ai/): Your team writes typed Python and wants agents that return validated data on any model.
- [Mastra](https://digidle.com/tools/mastra/): Your stack is TypeScript and you want agents, workflows, evals and tracing in one framework.
- [Google ADK](https://digidle.com/tools/google-adk/): You build on Gemini or Google Cloud, or need the same framework in Java, Go or Kotlin.
- [phpClaw](https://digidle.com/tools/phpclaw/): Your app runs on Laravel, Symfony or WordPress and you'd rather not stand up a Python service.
- [guilds.run](https://digidle.com/tools/guilds-run/): You want a standing team of agents on your own machine and a full record of what each one did.

---
Source: Digidle Tools (https://digidle.com/). Full index: https://digidle.com/llms.txt