Split View: 오픈소스 커뮤니티 소식 모음: 2026년 3월 주목할 프로젝트들

Open Source Community Digest: Noteworthy Projects of March 2026

• Overview
• 1. The AI Agent Revolution
  • OpenClaw — Fastest-Growing Open Source Project
  • ByteDance DeerFlow 2.0 — Open-Source SuperAgent Harness
  • Alibaba Page Agent — Control Web UIs with Natural Language
  • NousResearch Hermes Agent — Self-Learning AI Agent
  • Nvidia NemoClaw — Enterprise AI Agent Platform
• 2. The Rise of Research Automation
  • Karpathy's autoresearch — 22,983 Stars in 3 Days
• 3. Multimodal and Video AI
  • Google Gemini Embedding 2 — Unified Multimodal Embeddings
  • LTX 2.3 by Lightricks — Open-Source 4K Video Generation
• 4. Developer Tooling
  • Biome — The Complete ESLint + Prettier Replacement
  • Ladybird Browser — A Truly Independent Web Browser
• 5. Security and OSINT
  • Shadowbroker — 15 Real-Time Data Sources in One OSINT Tool
• 6. Community Events
  • Google Summer of Code 2026 — 22nd Year, Apply by March 31
• 7. Trend Analysis: GitHub AI Report Q1 2026
• Quiz
• References

Overview

March 2026 has been a remarkable month for the open source ecosystem. AI agent frameworks are proliferating at an unprecedented pace, research automation tools are gaining traction, and multimodal models continue to push boundaries. According to GitHub's 2026 AI Report, 267 new AI models were released in Q1 2026 alone — the majority open-source or domain-specialized.

This digest covers the most compelling projects and news from this month.

1. The AI Agent Revolution

OpenClaw — Fastest-Growing Open Source Project

One of the most dramatic GitHub growth stories in recent memory: OpenClaw surged from 9,000 to 210,000+ stars within just a few weeks.

OpenClaw is a personal AI assistant that runs entirely on your own hardware — no cloud dependency required. It delivers powerful automation while keeping your data local and private. What sets it apart is seamless integration with the messaging platforms people already use every day.

Supported platforms:

• WhatsApp, Telegram, Slack, Discord
• Signal, iMessage, Microsoft Teams

Key use cases:

• Developer workflow automation
• Personal productivity management
• Web scraping and data collection

# OpenClaw local setup using Ollama
git clone https://github.com/openclaw/openclaw
cd openclaw
ollama pull llama3.2
python setup.py --platform slack,telegram
python run.py --model llama3.2 --port 8080

The recipe for OpenClaw's success is straightforward: privacy-conscious users want to run AI on their own machines, connected to the communication tools they already rely on. The project delivers that in a polished, accessible package.

ByteDance DeerFlow 2.0 — Open-Source SuperAgent Harness

ByteDance released DeerFlow 2.0, described as an open-source "SuperAgent harness" capable of handling research, code generation, and content creation within a unified pipeline.

Core capabilities:

• Safe code execution inside sandboxed environments
• Long-term context retention via memory systems
• Integration with diverse external tools and APIs
• Skills library and sub-agent orchestration

# DeerFlow 2.0 usage example
from deerflow import SuperAgent, Sandbox

agent = SuperAgent(
    model="deepseek-r2",
    sandbox=Sandbox(type="docker"),
    memory=True,
    skills=["web_search", "code_exec", "doc_gen"]
)

result = agent.run(
    task="Research the latest open source trends and write a summary report",
    output_format="markdown"
)

DeerFlow 2.0 represents ByteDance's bet that the future of AI lies in composable, auditable agent pipelines rather than monolithic models.

Alibaba Page Agent — Control Web UIs with Natural Language

Alibaba's Page Agent is a JavaScript-based in-page GUI agent that takes a unique approach: instead of relying on screenshots (computationally expensive and fragile), it uses text-based DOM analysis to understand and interact with web interfaces.

// Page Agent usage example
import { PageAgent } from '@alibaba/page-agent'

const agent = new PageAgent({ model: 'qwen-max' })

// Control web interfaces with natural language
await agent.execute('Type "open source AI agents" in the search bar and click the first result')
await agent.execute('Fill out the contact form and submit it')

Being text-based makes it dramatically lighter than screenshot-based alternatives, while maintaining high accuracy on structured web content.

NousResearch Hermes Agent — Self-Learning AI Agent

Hermes Agent from NousResearch introduces a compelling concept: an AI agent that learns from repetition. As the agent encounters recurring tasks, it automatically identifies patterns, encodes them as "skills," and reuses those skills in future similar situations.

No manual programming of new routines is required — the agent optimizes its own behavior over time. This represents a meaningful step toward AI agents as genuine learning companions rather than static tools.

Nvidia NemoClaw — Enterprise AI Agent Platform

Nvidia announced NemoClaw, an open-source AI agent platform purpose-built for enterprise environments. It includes agentic tooling with security, compliance, and scalability in mind, tightly integrated with Nvidia's GPU infrastructure for high-performance inference.

2. The Rise of Research Automation

Karpathy's autoresearch — 22,983 Stars in 3 Days

Andrej Karpathy published autoresearch, which reached 22,983 GitHub stars within just three days. It also scored 198 points on Hacker News, demonstrating strong community interest.

Given a research query, the tool automatically performs web searches, collects relevant papers, generates summaries, and surfaces actionable insights. It is widely credited with sparking the current research automation movement.

# autoresearch usage example
pip install autoresearch
autoresearch --query "transformer architecture improvements 2025-2026" \
             --sources arxiv,semantic_scholar,web \
             --output report.md \
             --depth 3

Both academia and industry are taking notice, and a wave of similar tools is already emerging in its wake.

3. Multimodal and Video AI

Google Gemini Embedding 2 — Unified Multimodal Embeddings

Released on March 15, Gemini Embedding 2 is Google's unified multimodal embedding model that represents text, images, video, audio, and documents all within a single shared embedding space.

Previously, developers needed separate embedding models for each modality. Gemini Embedding 2 collapses this into a single model, significantly simplifying multimodal search, recommendation systems, and RAG (Retrieval-Augmented Generation) pipelines.

LTX 2.3 by Lightricks — Open-Source 4K Video Generation

Lightricks released LTX 2.3, a 22-billion-parameter open-source video generation model that sets a new quality bar for the field.

Key specifications:

• Parameters: 22B
• Resolution: Native 4K
• Frame rate: 50 FPS
• Audio: Synchronized audio generation
• Portrait mode: 1080x1920 (optimized for mobile content)

This is a landmark release for open-source video generation — matching or exceeding proprietary offerings in key metrics while being fully open.

4. Developer Tooling

Biome — The Complete ESLint + Prettier Replacement

Biome consolidates two essential JavaScript/TypeScript tools — ESLint (linting) and Prettier (formatting) — into a single, blazing-fast toolchain.

Performance comparison:

// biome.json configuration example
{
  "$schema": "https://biomejs.dev/schemas/1.9.0/schema.json",
  "linter": {
    "enabled": true,
    "rules": {
      "recommended": true
    }
  },
  "formatter": {
    "enabled": true,
    "indentStyle": "space",
    "indentWidth": 2
  }
}

# Lint + format in a single command
npx @biomejs/biome check --apply ./src

Written in Rust, Biome achieves its performance advantage by operating at native speed rather than through Node.js. For large codebases, the difference is dramatic.

Ladybird Browser — A Truly Independent Web Browser

Ladybird is an open-source web browser built entirely from scratch — no Chromium, no WebKit, no Gecko. Every component is original code, developed with a focus on performance, security, and privacy.

Ladybird originated from the SerenityOS project and has grown into a standalone effort to create a genuine alternative to the current browser duopoly. While still in active development, it represents one of the most ambitious open-source engineering projects of this era.

5. Security and OSINT

Shadowbroker — 15 Real-Time Data Sources in One OSINT Tool

Shadowbroker is an OSINT (Open-Source Intelligence) tool that integrates 15 real-time data sources into a single interface. It scored an impressive 304 points on Hacker News.

Capabilities include:

• Corporate jet and private aircraft tracking
• Satellite orbit data and analysis
• Maritime vessel tracking
• Corporate infrastructure mapping

It is particularly valuable for security researchers, investigative journalists, and corporate intelligence teams. As with all OSINT tools, ethical and lawful use is essential.

6. Community Events

Google Summer of Code 2026 — 22nd Year, Apply by March 31

Google Summer of Code (GSoC) 2026 marks its 22nd year.

• Participating organizations: 185 open source communities
• Application deadline: March 31, 2026
• Eligibility: Anyone 18 or older interested in contributing to open source

GSoC provides developers worldwide with a structured, paid opportunity to contribute meaningfully to open source projects under the guidance of experienced mentors. If you've been considering participating, the deadline is approaching.

7. Trend Analysis: GitHub AI Report Q1 2026

GitHub's 2026 AI Report paints a clear picture of where the open source world is heading.

Key figures:

• New AI models in Q1 2026: 267
• Share that are open-source or specialized: majority

Three leading trends:

1. Local inference: Demand for running AI on personal hardware, without cloud dependency, is growing rapidly — OpenClaw being the most visible example.

2. Workflow automation: Tools that replace repetitive tasks with AI-powered pipelines are proliferating across every domain.

3. Agent frameworks: Multi-agent systems that orchestrate several AI models and external tools are becoming the dominant architectural pattern.

The data confirms what practitioners have been sensing: open source AI is no longer experimental. It is production-ready and deeply embedded in how developers work.

Quiz

References

• OpenClaw GitHub Repository
• karpathy/autoresearch
• ByteDance DeerFlow 2.0
• Alibaba Page Agent
• NousResearch Hermes Agent
• Biome — Toolchain for the Web
• Ladybird Browser
• Google Summer of Code 2026
• GitHub AI Report Q1 2026
• Google Gemini Embedding 2
• LTX 2.3 by Lightricks