Starts Nov 23Live Cohort 3 of Vibe Coding & AI Agents: 3 weeks, 9 live hands-on sessionsCohort 3 (Nov 23) →
ReleaseCode / SDK · ActionableX: @replay_0x→ original

21 Free Open-Source Tools That Make Up an AI Agent Stack

The author of the @replay_0x channel has compiled 21 open-source repositories that can make up a full stack for AI agents, from orchestration to pre-release checks. The list follows the chain: orchestration, tools, browser, memory, observability, and safe code execution. Entries include LangGraph, CrewAI, MCP, browser-use, mem0, Langfuse, and E2B. For teams, it works as a ready-made map: one candidate per layer instead of searching from scratch.

AI-processed from X: @replay_0x; edited by Hamidun News
21 Free Open-Source Tools That Make Up an AI Agent Stack
Source: X: @replay_0x. Collage: Hamidun News.
Listen to article
Article Fact Sheet & Key Figures
ParameterDetails
Primary sourceX: @replay_0x ↗
Editor-in-Chief
OrchestrationLangGraph, CrewAI, AutoGen, OpenAI Agents SDK, Mastra
Memorymem0, Letta, and Graphiti, time-aware knowledge graphs
Before releasepromptfoo for tests, Langfuse for traces

When a team decides to "build an agent," the first week usually goes not to the product but to choosing tools: what manages the steps, how to connect services, where to store memory, and how to figure out why the agent suddenly answered with nonsense. A roundup of 21 open-source repositories, posted by the author of the @replay_0x channel, tries to solve this problem. It is laid out as a chain: define the task, organize the agents' work, connect tools, give them a browser, add memory, and ship to production. It is not a ranking or a benchmark but a map of layers, though it is useful as a starting point.

Layer 1.

Orchestration: who does the steps, and in what order

There are eight entries here, and many of them compete with one another.

Claude Agent SDK is the framework that Claude Code runs on, packaged as a library. LangGraph describes an agent as a graph with state: each step is a node, and the transitions between them are defined explicitly. CrewAI builds a "crew" of agents, each with its own role, like employees in a small department.

AutoGen from Microsoft bets on agents that message each other. OpenAI Agents SDK focuses on handing a task off from one agent to another. Mastra is the same idea for people who write in TypeScript.

Pydantic AI demands a typed, validated response from the agent: instead of free-form text, you get a structure you can pass straight into code. smolagents from Hugging Face offers agents that reason directly in code.

Layer 2.

Tools, browser, and memory

To do anything useful, an agent needs hands. MCP Servers is a set of ready-made servers built on the Model Context Protocol, the idea being to connect any tool to any agent in the same way. FastMCP lets you write your own MCP server for an internal service in a few minutes.

If a task has no API, the browser is what's left. browser-use lets agents click through real websites. Stagehand reduces working with a page to the commands act() and extract(): perform an action and pull out data. Playwright MCP controls the browser through the same MCP protocol.

Memory solves an everyday problem: an agent should not have to get to know you from scratch every time. mem0 stores long-term memory. Letta builds stateful agents that remember past sessions. Graphiti stores knowledge as a graph and tracks time, meaning it understands that a customer's address in March and in November may differ.

Layer 3.

Building, observability, and security

The last group is what you need to take a prototype to release. n8n is automation with more than 400 integrations, convenient when the agent is embedded in ordinary business processes. Dify lets you build LLM applications without code, a fast route for product managers and prototypes. Langfuse shows traces and request costs and helps debug chains. promptfoo runs quality evaluations and red-team attacks before release. E2B provides secure sandboxes where you can run agent-generated code without risking your own server.

What to do with this, and where the catch is

The practical way to use the list is to pick one tool per layer rather than adopting everything at once. For example, a typical internal agent needs only an orchestrator like LangGraph, a couple of MCP servers, Langfuse for observability, and promptfoo for pre-release checks.

There are also honest limitations. First, "100% free" applies to the code. Model tokens, hosting, browser sessions, and sandboxes cost money. Second, the orchestrators overlap: LangGraph, CrewAI, AutoGen, and OpenAI Agents SDK solve similar problems, and there is no reason to install them together. Third, a social media roundup says nothing about the maturity of the projects, their licenses, or the quality of support; you need to check these yourself before adopting anything. Finally, there is no data on how these tools work together: you will have to test the integrations on your own tasks.

Still, as an architecture checklist, the list works well: it is a reminder that an agent is not just a model. Without memory, observability, tests, and code isolation, it remains a demo. If your project lacks even the last three layers, start with them.

Story

Key Questions Answered

What are the key details on Layer 1. Orchestration: who does the steps, and in what order?

There are eight entries here, and many of them compete with one another. Claude Agent SDK is the framework that Claude Code runs on, packaged as a library. LangGraph describes an agent as a graph with state: each step is a node, and the transitions between them are defined explicitly. CrewAI builds a "crew" of agents, each with its own role, like employees in a small department. AutoGen from Microsoft bets on agents that message each other. OpenAI Agents SDK focuses on handing a task off from one agent to another.

What are the key details on Layer 2. Tools, browser, and memory?

To do anything useful, an agent needs hands. MCP Servers is a set of ready-made servers built on the Model Context Protocol, the idea being to connect any tool to any agent in the same way. FastMCP lets you write your own MCP server for an internal service in a few minutes. If a task has no API, the browser is what's left. browser-use lets agents click through real websites. Stagehand reduces working with a page to the commands act() and extract(): perform an action and pull out data.

What are the key details on Layer 3. Building, observability, and security?

The last group is what you need to take a prototype to release. n8n is automation with more than 400 integrations, convenient when the agent is embedded in ordinary business processes. Dify lets you build LLM applications without code, a fast route for product managers and prototypes. Langfuse shows traces and request costs and helps debug chains. promptfoo runs quality evaluations and red-team attacks before release.

ZK
Zhemal Khamidun — CPO at Alpina Digital, product manager of AlpinaGPT. Building AI products since 2019. Hamidun News is a pet project that grew into a production aggregator with 6,000+ readers.
✦ Hands-On AI Engineering · Hamidun Academy

Build & Ship Production AI Agents Without Boilerplate Code

Start your free 72-hour trial at Hamidun Academy (academy.hamidun.com) — interactive lessons on Claude Code & Cursor, live Tech Radar, and production agent templates.

Start 72h Free Trial →Vibe-Coding Intensive
What do you think?
Loading comments…