01
Visual IDE for Multi-Agent Systems
I recently came across a project called Rowboat on Hacker News, described as the "Cursor for multi-agent systems."
After taking a look, I found it quite interesting.
Rowboat is an open-source project incubated by YC, currently boasting 12,000 stars.
Simply put, it is a visual IDE for building multi-agent systems, complete with a Copilot to assist in generating agents.
You don't need to write a single line of code; just throw in a natural language description, and the Copilot will build the multi-agent workflow for you.
After building, you can test it in AI-simulated scenarios. Once confirmed, you can connect it to MCP servers and various tools.
Under the hood, it uses OpenAI's Agents SDK and integrates with common services like Slack, Linear, Jira, GitHub, ElevenLabs, and Exa.
It supports both Python SDK and HTTP API methods, allowing integration into your own products.
If you want to build AI customer service, automated research, or internal workflows without starting from scratch with multi-agent frameworks, this project can save you a lot of time.
Open Source Address: https://github.com/rowboatlabs/rowboat02
Turning Coding Agents into True Teammates
Multica's approach is quite interesting: a combination of Linear and AI Agents.
Currently, when people use coding agents like Claude Code or Codex, they mostly copy and paste prompts back and forth in the terminal, and humans still need to monitor them after every task.
Multica aims to change this.
It turns agents into team members. You assign tasks on a board just like you would to a colleague; the agent executes them, reports progress, updates status, and even reports blockers if issues arise.
It has already garnered 14,700 stars and was updated in April.
Core capabilities include full lifecycle task management, real-time WebSocket progress streams, isolated workspaces, and support for hybrid local daemon and cloud runtime usage.
It is compatible with a wide range of CLIs including Claude Code, Codex, OpenCode, Gemini, and Cursor Agent.
Furthermore, solutions are automatically distilled into reusable skills, making the team stronger with usage.
If your team has started using AI agents for coding, this project can make collaboration orderly.
Open Source Address: https://github.com/multica-ai/multica03
AI Coding Engineering Discipline Pack
The author of this project is Addy Osmani, an engineering leader at Google Chrome and the author of the book "Learning JavaScript Design Patterns".
Agent Skills does something very direct: it encapsulates the development standards of senior engineers into skills that AI can execute directly.
It currently has 16,600 stars.
The core consists of 7 skills spanning the entire process: /spec for requirements, /plan for task breakdown, /build for incremental implementation, /test for verification, /review for quality gates, /code-simplify for simplification, and /ship for deployment.
It packages 20 skills, clearly divided by phase.
It basically covers the complete lifecycle from defining requirements to going live.
AI coding tends to take shortcuts and skip specifications; this project forces that big-tech engineer discipline onto the agent.
If you use Claude Code or similar tools for actual projects, this skill set can significantly improve delivery quality.
Open Source Address: https://github.com/addyosmani/agent-skills04
Workflow Engine for Reproducible AI Programming
Archon is an open-source project by coleam00, positioned as a harness builder for AI programming, currently with 18,400 stars.
It addresses a very painful problem.
Currently, when AI programming agents run, the results vary every time. For the same task, it might skip the planning phase today, forget to write tests tomorrow, and ignore PR templates the day after.
What Archon does is fix the process using YAML, making AI coding deterministic and reproducible.
Several design choices are quite thoughtful.
Each workflow run opens an independent git worktree, so running multiple tasks in parallel won't cause cross-contamination.
Nodes are composable; you can plug in deterministic bash scripts, tests, and AI planning or code generation nodes.
It comes with 17 default workflows, with templates for feature development, issue fixing, PR review, and refactoring.
It's not just a CLI tool; it can be triggered via CLI, Web UI, Slack, Telegram, Discord, and GitHub.
Teams just need to commit the YAML files under the .archon/workflows/ directory to the repository, and everyone aligns with the same processes.
It's similar to the role n8n plays in general automation; Archon wants to do the same for AI coding.
Open Source Address: https://github.com/coleam00/Archon05
AI Tutor Open Sourced by HKU
DeepTutor is an AI learning assistant open-sourced by the Data Intelligence Lab at the University of Hong Kong, currently holding 18,800 stars.
It features 5 major learning modes:
Tool-enhanced chat (RAG + Web + Code Execution), Deep Solve for breaking down complex problems with multi-agents, and Quiz Generation based on knowledge bases.
It also includes Deep Research for parallel agent deep dives and Math Animator specifically for math visualization.
Most interestingly, it created something called TutorBot. Each TutorBot has an independent workspace and persona, capable of autonomous operation, a step above simple chatbots.
There is also Co-Writer, a Markdown editor that treats AI as a first-class citizen, and Persistent Memory that shares the learner's profile across all functions.
Open Source Address: https://github.com/HKUDS/DeepTutor06
The CLAUDE.md That Makes Claude Code Smarter
After looking at this project, I feel it's quite top-tier.
The core of andrej-karpathy-skills is just a CLAUDE.md file. Inspired by Karpathy's complaints on Twitter about large models writing code, it has already amassed 50,000 stars.
What did Karpathy complain about?
Models making wrong assumptions and writing anyway without clarifying, over-designing code and APIs with useless abstractions, changing code they don't fully understand, and pretending to understand when confused instead of asking.
The author translated these observations into 4 principles:
Think before writing: Clarify assumptions, push back when necessary. Simplicity first: Write only necessary code, no speculative features. Surgical modifications: Touch only necessary code, match existing styles. Goal-driven execution: Set verification criteria before running.
There are two installation methods; the plugin route is recommended:
/plugin marketplace add forrestchang/andrej-karpathy-skills
/plugin install andrej-karpathy-skills@karpathy-skillsYou can also just download the CLAUDE.md file to the project root directory.
If you feel Claude Code has been a bit too presumptuous lately, try this file; it will behave much better.
Open Source Address: https://github.com/forrestchang/andrej-karpathy-skills07
Equipping Claude Code with Long-Term Memory
claude-mem solves the problem of Claude Code forgetting everything when starting a new session. It currently has 60,000 stars.
Its working principle is very direct: during a session, it automatically captures everything Claude does, uses Claude's agent-sdk for AI semantic compression, and injects the relevant context back when starting a new session next time.
You don't need to record manually or recall manually; the whole process is automatic.
The functionality is quite comprehensive.
Search uses skill-based natural language search with progressive disclosure. The token cost for each layer of memory retrieval is marked, so it won't silently burn your API fees.
It comes with a local Web Viewer running on localhost:37777, allowing you to view history directly in the browser.
Regarding privacy, you can use the <private> tag to wrap content you don't want to be remembered.
The underlying storage is SQLite plus Chroma vector database, all local.
It also supports Gemini CLI by the way.
Installation is done with one command:
npx claude-mem installIf you deal with Claude Code daily, this tool is almost a necessity. After installation, that feeling of picking up where you left off is quite comfortable.
Open Source Address: https://github.com/thedotmack/claude-mem08
Complete Package of Chinese Textbook PDFs
The ChinaTextbook project is quite hardcore; it collects PDFs of textbooks for elementary, middle, high school, and university, currently holding 69,700 stars.
The author's motivation is straightforward: to promote equity in compulsory education resources, allowing families without means to access genuine textbooks, and also facilitating overseas Chinese families in accessing Chinese educational resources for their children.
Content covers the People's Education Press edition, the "5-4" schooling system, and various other curricula. The math section spans from Grade 1 all the way to university, extremely comprehensive.
All files are PDF and completely free.
Some files exceeding 50MB are split into 35MB segments; the repository provides a merge tool.
The README specifically calls out resellers who take public resources and sell them for profit, drawing a clear line in the sand.
If you have children needing homework tutoring, or you want to brush up on some basic knowledge yourself, cloning this repository will serve you well for a long time.
Open Source Address: https://github.com/TapXWorld/ChinaTextbook09
God-Tier Tool for Converting Everything to Markdown
If you work on AI applications or RAG pipelines, you basically can't get around this project.
MarkItDown is an official Python tool from Microsoft, currently with 111,000 stars, definitely a top-tier project.
What it does is simple: convert various file formats into Markdown.
It supports a long list of formats: PDF, Word, PPT, Excel, images, audio, HTML, and even YouTube links can be dragged directly in.
Why convert to Markdown?
Because current LLMs are natively friendly to Markdown, offering the highest token utilization, and the structure is preserved completely—headings, lists, tables, and links all remain.
Several practical points: built-in LLM integration, images can use OpenAI models for descriptions, and audio can be transcribed. It connects with Azure Document Intelligence and supports third-party plugin extensions like markitdown-ocr.
Installation is one line:
pip install 'markitdown[all]'Both CLI and Python API are available:
markitdown path-to-file.pdf -o document.mdWhether feeding documents to Claude or GPT, building knowledge bases, or data cleaning, this tool should be in your common toolbox.
Open Source Address: https://github.com/microsoft/markitdown10
TTS Large Model Runnable on Edge Devices
VoxCPM is a Text-to-Speech large model open-sourced by ModelBest, and it has already secured 13,800 stars shortly after launch.
Simply put, this is a 2B parameter TTS model. Trained on over 2 million hours of multi-language speech data, it supports automatic recognition and switching for 30 languages, outputting audio at studio-quality 48kHz.
Two of its features are quite impressive.
One is Voice Design: you give it a text description, and it directly generates a voice tone matching the description without needing a reference audio clip.
The other is controllable voice cloning; after cloning, you can add style guidance to make the same voice express different emotions.
In terms of real-time streaming inference, the RTF can reach around 0.3 on a single RTX 4090, basically making it usable on the edge.
If you are working on podcasts, video dubbing, intelligent customer service, or audiobooks, you should pull this model down and try it.
Open Source Address: https://github.com/OpenBMB/VoxCPM