Open Projects
For engineers and researchers who build AI systems
Open-source AI projects that aren't products yet but are worth running and studying — agents, coding agents, infrastructure, and Korean open weights.
Personal & general agents
Open-source agents you run yourself — not products, but working systems worth studying and deploying. This is where agent architecture ideas ship first.
Self-hosted agent with persistent memory that writes its own reusable skills from experience, runs scheduled jobs, and connects to Telegram, Discord, Slack, and the terminal.
The built-in learning loop — it improves the longer you use it
Local-first personal AI assistant that runs on your own devices, supports every major LLM, and connects to 20+ messaging channels. Crossed 100K GitHub stars within its first week (Jan 2026).
The fastest-growing open agent project to date
Extensible open agent framework from Block that automates engineering tasks end-to-end, built natively on MCP for tool connections.
A large company running its internal agent fully in the open
Minimal agent library (~1,000 core lines) where agents write Python code to act instead of emitting JSON tool calls. The cleanest codebase to learn agent internals from.
Read the whole thing in an afternoon
Lets LLMs run code on your machine from natural language — a local, open counterpart to hosted code-execution sandboxes, with safeguards you control.
The 'computer as a tool' idea, running locally
Agent server built around the MemGPT research: agents with self-editing long-term memory that persist across sessions, exposed through a clean API.
Where agent memory research became runnable infrastructure
Coding agents
Open alternatives to Cursor and Copilot — study how autonomous software engineering actually works, or run them on your own models.
Platform for AI software development agents that write code, run commands, browse the web, and fix issues end-to-end. Formerly OpenDevin.
The reference open implementation of an autonomous dev agent
Terminal pair-programmer that edits code in your local git repo, works with Claude, GPT, and local models, and commits with sensible messages.
Proof that a great coding agent doesn't need an IDE
Open-source autonomous coding agent inside VS Code — creates and edits files, runs commands, uses the browser, with human approval at each step.
Bring-your-own-key alternative to closed coding agents
The academic system behind SWE-bench: turns a language model into a software engineering agent that fixes real GitHub issues, with a carefully designed agent-computer interface.
Where the agent-computer interface idea was formalized
Open-source terminal coding agent in the Claude Code mold — provider-agnostic, scriptable, with a polished TUI. Run it on Claude, GPT, or open-weight models.
The strongest open answer to closed terminal agents
Open IDE assistant for VS Code and JetBrains — chat, autocomplete, and edit with any model, plus shareable team configs. The bring-your-own-model default for IDE AI.
IDE AI without vendor lock-in
Agent infrastructure
The plumbing serious agent systems are built on — browser control and model serving.
Makes websites accessible to AI agents — extracts interactive elements and lets any LLM click, type, and navigate. The de-facto standard library for browser agents.
Powering most open web-agent demos you've seen
High-throughput LLM serving engine with PagedAttention — the default way to serve open-weight models (including Korean models like EXAONE and Solar) in production.
If you self-host a model, you probably run this
One-command local model runtime — pull and run open-weight models (EXAONE and HyperCLOVA X SEED included) on a laptop, with an OpenAI-compatible local API.
How most developers first run a local model
Serving engine with RadixAttention prefix caching and a structured-generation language — the main open rival to vLLM, strong on agentic workloads with repeated prefixes.
The other engine serious serving teams benchmark against
Unified gateway that speaks the OpenAI format to 100+ providers — one proxy for keys, budgets, fallbacks, and routing across Claude, GPT, Gemini, and Korean APIs alike.
The de-facto open model gateway
Korean open weights
Korea's sovereign model families ship open weights on Hugging Face. Download, fine-tune, and serve them with vLLM — no API contract needed.
LG's frontier Korean-English model family, including reasoning and multimodal variants, under one of Korea's five MSIT sovereign-AI mandates.
The most-downloaded Korean open-weight family
Naver's open-weight SEED lineup — efficient Korean-first models for deployment, from small on-device sizes to the SEED Think reasoning line.
Korean-first tokenizer efficiency you can actually measure
Upstage's Solar open models — the first Korean family recognized as frontier-class by Artificial Analysis, with open releases alongside the commercial API.
Frontier-class recognition with open releases
Kakao's Kanana model family with open-weight releases — compute-efficient bilingual models built for KakaoTalk-scale consumer deployment.
Consumer-platform Korean models in the open
Multi-agent & orchestration frameworks
Frameworks for composing agents into teams and pipelines — the layer where single-agent demos become systems.
Microsoft's multi-agent framework — agents that converse, critique, and delegate to each other, with the AgentChat API and no-code Studio on top.
The most-cited multi-agent conversation pattern
Role-based agent teams — define a crew with roles, goals, and tools, and let them execute processes together. The fastest on-ramp from idea to working multi-agent flow.
Role-play as an engineering primitive
Google's open framework behind Vertex AI agents — hierarchical agent composition, built-in evaluation, and deployment paths from laptop to cloud.
A hyperscaler's production agent framework, in the open
Agent framework from the Pydantic team — type-safe tools and structured outputs with the validation rigor Python backends already trust.
Type safety as the agent reliability story
RAG & document intelligence
Retrieval and document parsing — where enterprise AI actually lives, and where Korean document formats need the most work.
End-to-end RAG engine with deep document understanding — layout-aware chunking, citations grounded in the source, and a full web UI out of the box.
RAG that takes document layout seriously
Graph-enhanced retrieval that builds entity-relation graphs from documents and mixes graph and vector search — stronger multi-hop answers than flat chunking.
GraphRAG ideas in a codebase you can actually read
Document conversion for AI pipelines — PDF, Office, and scanned files into clean structured Markdown/JSON, with table structure and reading order preserved.
The parsing layer most RAG stacks were missing
High-fidelity PDF extraction tuned for scientific and complex documents — formulas, tables, and multi-column layouts into Markdown for model consumption.
What to reach for when Docling isn't enough
Korean open-source AI tools
The Korean-language layer — morphology, NLP, learning resources, and the HWP document format our challenge is trying to crack open.
Fast modern Korean morphological analyzer in C++ with Python bindings — the practical choice for tokenizing Korean text in production pipelines today.
The workhorse of modern Korean text processing
The classic Korean NLP toolkit that unified a generation of morphological analyzers behind one Python API — still the first import in most Korean NLP notebooks.
Where Korean NLP standardized
The most complete Korean-language LLM engineering curriculum on GitHub — hands-on LangChain/RAG/agent notebooks that much of Korea's AI developer community learned from.
Korea's de-facto open LLM textbook
Open-source HWP viewer and parser in TypeScript — one of the few serious open attacks on Korea's closed document format, and directly relevant to our HWP challenge.
Proof the HWP format can be opened
Robotics & embodied AI
Wave 5 in the open: the projects making robot learning reproducible outside big labs — where Korea's manufacturing edge meets open software.
Open robot-learning stack — datasets, imitation and RL policies, and low-cost robot arm designs that made real-robot training accessible to individuals.
Robot learning's Hugging Face moment
Generative physics engine for robotics — ultra-fast parallel simulation designed to train embodied agents at scales physical hardware can't match.
Simulation speed as the new training data
Open vision-language-action model for robot manipulation — the reference open checkpoint for turning camera frames and instructions into robot actions.
The open baseline every VLA paper compares against