来这儿
看帖
0 FEEDBACK
huaren wxcity hn newmitbbs bedtime bili
new
jiandan
new
hdeal github
Prev day Apr 01 Wed Next day Today
agent-lightning

The absolute trainer to light up AI agents.

Agent-lightning-banner

Agent Lightning⚡

Unit Tests
Documentation
PyPI version
License
Ask DeepWiki
Discord

The absolute trainer to light up AI agents.

Join our Discord community to connect with other users and contributors.

⚡ Core Features

  • Turn your agent into an optimizable beast with ZERO CODE CHANGE (almost)! 💤
  • Build with ANY agent framework (LangChain, OpenAI Agent SDK, AutoGen, CrewAI, Microsoft Agent Framework...); or even WITHOUT agent framework (Python OpenAI). You name it! 🤖
  • Selectively optimize one or more agents in a multi-agent system. 🎯
  • Embraces Algorithms like Reinforcement Learning, Automatic Prompt Optimization, Supervised Fine-tuning and more. 🤗

Read more on our documentation website.

Agent-Lightning Core Quickstart

⚡ Installation

pip install agentlightning

For the latest nightly build (cutting-edge features), you can install from Test PyPI:

pip install --upgrade --index-url https://test.pypi.org/simple/ --extra-index-url https://pypi.org/simple/ --pre agentlightning

Please refer to our installation guide for more details.

To start using Agent-lightning, check out our documentation and examples.

⚡ Articles

⚡ Community Projects

⚡ Architecture

Agent Lightning keeps the moving parts to a minimum so you can focus on your idea, not the plumbing. Your agent continues to run as usual; you can still use any agent framework you like; you drop in the lightweight agl.emit_xxx() helper, or let the tracer collect every prompt, tool call, and reward. Those events become structured spans that flow into the LightningStore, a central hub that keeps tasks, resources, and traces in sync.

On the other side of the store sits the algorithm you choose, or write yourself. The algorithm reads spans, learns from them, and posts updated resources such as refined prompt templates or new policy weights. The Trainer ties it all together: it streams datasets to runners, ferries resources between the store and the algorithm, and updates the inference engine when improvements land. You can either stop there, or simply let the same loop keep turning.

No rewrites, no lock-in, just a clear path from first rollout to steady improvement.

Agent-lightning Architecture

⚡ CI Status

Workflow Status
CPU Tests tests workflow status
Full Tests tests summary workflow status
UI Tests UI Tests
Examples Integration examples summary workflow status
Latest Dependency Compatibility latest summary workflow status
Legacy Examples Compatibility compat summary workflow status

⚡ Citation

If you find Agent Lightning useful in your research or projects, please cite our paper:

@misc{luo2025agentlightningtrainai,
      title={Agent Lightning: Train ANY AI Agents with Reinforcement Learning},
      author={Xufang Luo and Yuge Zhang and Zhiyuan He and Zilong Wang and Siyun Zhao and Dongsheng Li and Luna K. Qiu and Yuqing Yang},
      year={2025},
      eprint={2508.03680},
      archivePrefix={arXiv},
      primaryClass={cs.AI},
      url={https://arxiv.org/abs/2508.03680},
}

⚡ Contributing

This project welcomes contributions and suggestions. Start by reading the Contributing Guide for recommended contribution points, environment setup, branching conventions, and pull request expectations. Most contributions require you to agree to a Contributor License Agreement (CLA) declaring that you have the right to, and actually do, grant us the rights to use your contribution. For details, visit https://cla.opensource.microsoft.com.

When you submit a pull request, a CLA bot will automatically determine whether you need to provide a CLA and decorate the PR appropriately (e.g., status check, comment). Simply follow the instructions provided by the bot. You will only need to do this once across all repos using our CLA.

This project has adopted the Microsoft Open Source Code of Conduct. For more information see the Code of Conduct FAQ or contact opencode@microsoft.com with any additional questions or comments.

⚡ Trademarks

This project may contain trademarks or logos for projects, products, or services. Authorized use of Microsoft trademarks or logos is subject to and must follow Microsoft's Trademark & Brand Guidelines. Use of Microsoft trademarks or logos in modified versions of this project must not cause confusion or imply Microsoft sponsorship. Any use of third-party trademarks or logos are subject to those third-party's policies.

⚡ Responsible AI

This project has been evaluated and certified to comply with the Microsoft Responsible AI Standard. The team will continue to monitor and maintain the repository, addressing any severe issues, including potential harms, if they arise.

⚡ License

This project is licensed under the MIT License. See the LICENSE file for details.

▼ Show full
coding-interview-university

A complete computer science study plan to become a software engineer.

Coding Interview University

I originally created this as a short to-do list of study topics for becoming a software engineer,
but it grew to the large list you see today. After going through this study plan, I got hired
as a Software Development Engineer at Amazon!
You probably won't have to study as much as I did. Anyway, everything you need is here.

I studied about 8-12 hours a day, for several months. This is my story: Why I studied full-time for 8 months for a Google interview

Please Note: You won't need to study as much as I did. I wasted a lot of time on things I didn't need to know. More info about that is below. I'll help you get there without wasting your precious time.

The items listed here will prepare you well for a technical interview at just about any software company,
including the giants: Amazon, Facebook, Google, and Microsoft.

Best of luck to you!

Translations:
Translations in progress:

What is it?

Coding at the whiteboard - from HBO's Silicon Valley

This is my multi-month study plan for becoming a software engineer for a large company.

Required:

  • A little experience with coding (variables, loops, methods/functions, etc)
  • Patience
  • Time

Note this is a study plan for software engineering, not frontend engineering or full-stack development. There are really
super roadmaps and coursework for those career paths elsewhere (see https://roadmap.sh/ for more info).

There is a lot to learn in a university Computer Science program, but only knowing about 75% is good enough for an interview, so that's what I cover here.
For a complete CS self-taught program, the resources for my study plan have been included in Kamran Ahmed's Computer Science Roadmap: https://roadmap.sh/computer-science

Table of Contents

The Study Plan

Topics of Study

Getting the Job

---------------- Everything below this point is optional ----------------

Optional Extra Topics & Resources

Why use it?

If you want to work as a software engineer for a large company, these are the things you have to know.

If you missed out on getting a degree in computer science, like I did, this will catch you up and save four years of your life.

When I started this project, I didn't know a stack from a heap, didn't know Big-O anything, or anything about trees, or how to
traverse a graph. If I had to code a sorting algorithm, I can tell ya it would have been terrible.
Every data structure I had ever used was built into the language, and I didn't know how they worked
under the hood at all. I never had to manage memory unless a process I was running would give an "out of
memory" error, and then I'd have to find a workaround. I used a few multidimensional arrays in my life and
thousands of associative arrays, but I never created data structures from scratch.

It's a long plan. It may take you months. If you are familiar with a lot of this already it will take you a lot less time.

⬆ back to top

How to use it

Everything below is an outline, and you should tackle the items in order from top to bottom.

I'm using GitHub's special markdown flavor, including tasks lists to track progress.

If you don't want to use git

On this page, click the Code button near the top, then click "Download ZIP". Unzip the file and you can work with the text files.

If you're open in a code editor that understands markdown, you'll see everything formatted nicely.

How to download the repo as a zip file

If you're comfortable with git

Create a new branch so you can check items like this, just put an x in the brackets: [x]

  1. Fork the GitHub repo: https://github.com/jwasham/coding-interview-university by clicking on the Fork button.

    Fork the GitHub repo

  2. Clone to your local repo:

    git clone https://github.com/<YOUR_GITHUB_USERNAME>/coding-interview-university.git
cd coding-interview-university
git remote add upstream https://github.com/jwasham/coding-interview-university.git
git remote set-url --push upstream DISABLE  # so that you don't push your personal progress back to the original repo

  3. Mark all boxes with X after you completed your changes:

    git commit -am "Marked personal progress"
git pull upstream main  # keep your fork up-to-date with changes from the original repo

git push # just pushes to your fork

⬆ back to top

Don't feel you aren't smart enough

⬆ back to top

A Note About Video Resources

Some videos are available only by enrolling in a Coursera or EdX class. These are called MOOCs.
Sometimes the classes are not in session so you have to wait a couple of months, so you have no access.

It would be great to replace the online course resources with free and always-available public sources,
such as YouTube videos (preferably university lectures), so that you people can study these anytime,
not just when a specific online course is in session.

⬆ back to top

Choose a Programming Language

You'll need to choose a programming language for the coding interviews you do,
but you'll also need to find a language that you can use to study computer science concepts.

Preferably the language would be the same, so that you only need to be proficient in one.

For this Study Plan

When I did the study plan, I used 2 languages for most of it: C and Python

  • C: Very low level. Allows you to deal with pointers and memory allocation/deallocation, so you feel the data structures
    and algorithms in your bones. In higher-level languages like Python or Java, these are hidden from you. In day-to-day work, that's terrific,
    but when you're learning how these low-level data structures are built, it's great to feel close to the metal.
    • C is everywhere. You'll see examples in books, lectures, videos, everywhere while you're studying.
    • The C Programming Language, 2nd Edition
      • This is a short book, but it will give you a great handle on the C language and if you practice it a little
        you'll quickly get proficient. Understanding C helps you understand how programs and memory work.
      • You don't need to go super deep in the book (or even finish it). Just get to where you're comfortable reading and writing in C.
  • Python: Modern and very expressive, I learned it because it's just super useful and also allows me to write less code in an interview.

This is my preference. You do what you like, of course.

You may not need it, but here are some sites for learning a new language:

For your Coding Interview

You can use a language you are comfortable in to do the coding part of the interview, but for large companies, these are solid choices:

  • C++
  • Java
  • Python

You could also use these, but read around first. There may be caveats:

  • JavaScript
  • Ruby

Here is an article I wrote about choosing a language for the interview:
Pick One Language for the Coding Interview.
This is the original article my post was based on: Choosing a Programming Language for Interviews

You need to be very comfortable in the language and be knowledgeable.

Read more about choices:

See language-specific resources here

⬆ back to top

Books for Data Structures and Algorithms

This book will form your foundation for computer science.

Just choose one, in a language that you will be comfortable with. You'll be doing a lot of reading and coding.

Python

  • Coding Interview Patterns: Nail Your Next Coding Interview (Main Recommendation)
    • An insider’s perspective on what interviewers are truly looking for and why.
    • 101 real coding interview problems with detailed solutions.
    • Intuitive explanations that guide you through each problem as if you were solving it in a live interview.
    • 1000+ diagrams to illustrate key concepts and patterns.

C

Java

Your choice:

C++

Your choice:

⬆ back to top

Interview Prep Books

Here are some recommended books to supplement your learning.

If you have tons of extra time:

Choose one:

⬆ back to top

Don't Make My Mistakes

This list grew over many months, and yes, it got out of hand.

Here are some mistakes I made so you'll have a better experience. And you'll save months of time.

1. You Won't Remember it All

I watched hours of videos and took copious notes, and months later there was much I didn't remember. I spent 3 days going
through my notes and making flashcards, so I could review. I didn't need all of that knowledge.

Please, read so you won't make my mistakes:

Retaining Computer Science Knowledge.

2. Use Flashcards

To solve the problem, I made a little flashcard site where I could add flashcards of 2 types: general and code.
Each card has a different formatting. I made a mobile-first website, so I could review on my phone or tablet, wherever I am.

Make your own for free:

I DON'T RECOMMEND using my flashcards. There are too many and most of them are trivia that you don't need.

But if you don't want to listen to me, here you go:

Keep in mind I went overboard and have cards covering everything from assembly language and Python trivia to machine learning and statistics.
It's way too much for what's required.

Note on flashcards: The first time you recognize you know the answer, don't mark it as known. You have to see the
same card and answer it several times correctly before you really know it. Repetition will put that knowledge deeper in
your brain.

An alternative to using my flashcard site is Anki, which has been recommended to me numerous times.
It uses a repetition system to help you remember. It's user-friendly, available on all platforms, and has a cloud sync system.
It costs $25 on iOS but is free on other platforms.

My flashcard database in Anki format: https://ankiweb.net/shared/info/25173560 (thanks @xiewenya).

Some students have mentioned formatting issues with white space that can be fixed by doing the following: open the deck, edit the card, click cards, select the "styling" radio button, and add the member "white-space: pre;" to the card class.

3. Do Coding Interview Questions While You're Learning

THIS IS VERY IMPORTANT.

Start doing coding interview questions while you're learning data structures and algorithms.

You need to apply what you're learning to solve problems, or you'll forget. I made this mistake.

Once you've learned a topic, and feel somewhat comfortable with it, for example, linked lists:

  1. Open one of the coding interview books (or coding problem websites, listed below)
  2. Do 2 or 3 questions regarding linked lists.
  3. Move on to the next learning topic.
  4. Later, go back and do another 2 or 3 linked list problems.
  5. Do this with each new topic you learn.

Keep doing problems while you're learning all this stuff, not after.

You're not being hired for knowledge, but how you apply the knowledge.

There are many resources for this, listed below. Keep going.

4. Focus

There are a lot of distractions that can take up valuable time. Focus and concentration are hard. Turn on some music
without lyrics and you'll be able to focus pretty well.

⬆ back to top

What you won't see covered

These are prevalent technologies but not part of this study plan:

  • Javascript
  • HTML, CSS, and other front-end technologies
  • SQL

⬆ back to top

The Daily Plan

This course goes over a lot of subjects. Each will probably take you a few days, or maybe even a week or more. It depends on your schedule.

Each day, take the next subject in the list, watch some videos about that subject, and then write an implementation
of that data structure or algorithm in the language you chose for this course.

You can see my code here:

You don't need to memorize every algorithm. You just need to be able to understand it enough to be able to write your own implementation.

⬆ back to top

Coding Question Practice

Why is this here? I'm not ready to interview.

Then go back and read this.

Why you need to practice doing programming problems:

  • Problem recognition, and where the right data structures and algorithms fit in
  • Gathering requirements for the problem
  • Talking your way through the problem like you will in the interview
  • Coding on a whiteboard or paper, not a computer
  • Coming up with time and space complexity for your solutions (see Big-O below)
  • Testing your solutions

There is a great intro for methodical, communicative problem-solving in an interview. You'll get this from the programming
interview books, too, but I found this outstanding:
Algorithm design canvas

Write code on a whiteboard or paper, not a computer. Test with some sample inputs. Then type it and test it out on a computer.

If you don't have a whiteboard at home, pick up a large drawing pad from an art store. You can sit on the couch and practice.
This is my "sofa whiteboard". I added the pen in the photo just for scale. If you use a pen, you'll wish you could erase.
Gets messy quickly. I use a pencil and eraser.

my sofa whiteboard

Coding question practice is not about memorizing answers to programming problems.

⬆ back to top

Coding Problems

Don't forget your key coding interview books here.

Solving Problems:

Coding Interview Question Videos:

Challenge/Practice sites:

  • LeetCode
    • My favorite coding problem site. It's worth the subscription money for the 1-2 months you'll likely be preparing.
    • See Nick White and FisherCoder Videos above for code walk-throughs.
  • HackerRank
  • TopCoder
  • Codeforces
  • Codility
  • Geeks for Geeks
  • AlgoExpert
    • Created by Google engineers, this is also an excellent resource to hone your skills.
  • Project Euler
    • very math-focused, and not really suited for coding interviews

⬆ back to top

Let's Get Started

Alright, enough talk, let's learn!

But don't forget to do coding problems from above while you learn!

Algorithmic complexity / Big-O / Asymptotic analysis

Well, that's about enough of that.

When you go through "Cracking the Coding Interview", there is a chapter on this, and at the end there is a quiz to see
if you can identify the runtime complexity of different algorithms. It's a super review and test.

⬆ back to top

Data Structures

⬆ back to top

More Knowledge

⬆ back to top

Trees

⬆ back to top

Sorting

As a summary, here is a visual representation of 15 sorting algorithms.
If you need more detail on this subject, see the "Sorting" section in Additional Detail on Some Subjects

⬆ back to top

Graphs

Graphs can be used to represent many problems in computer science, so this section is long, like trees and sorting.

⬆ back to top

Even More Knowledge

⬆ back to top

Final Review

This section will have shorter videos that you can watch pretty quickly to review most of the important concepts.
It's nice if you want a refresher often.

⬆ back to top

Update Your Resume

⬆ back to top

Interview Process & General Interview Prep

Mock Interviews:

⬆ back to top

Be thinking of for when the interview comes

Think of about 20 interview questions you'll get, along with the lines of the items below. Have at least one answer for each.
Have a story, not just data, about something you accomplished.

  • Why do you want this job?
  • What's a tough problem you've solved?
  • Biggest challenges faced?
  • Best/worst designs seen?
  • Ideas for improving an existing product
  • How do you work best, as an individual and as part of a team?
  • Which of your skills or experiences would be assets in the role and why?
  • What did you most enjoy at [job x / project y]?
  • What was the biggest challenge you faced at [job x / project y]?
  • What was the hardest bug you faced at [job x / project y]?
  • What did you learn at [job x / project y]?
  • What would you have done better at [job x / project y]?

⬆ back to top

Have questions for the interviewer

Some of mine (I already may know the answers, but want their opinion or team perspective):

  • How large is your team?
  • What does your dev cycle look like? Do you do waterfall/sprints/agile?
  • Are rushes to deadlines common? Or is there flexibility?
  • How are decisions made in your team?
  • How many meetings do you have per week?
  • Do you feel your work environment helps you concentrate?
  • What are you working on?
  • What do you like about it?
  • What is the work life like?
  • How is the work/life balance?

⬆ back to top

Once You've Got The Job

Congratulations!

Keep learning.

You're never really done.

*****************************************************************************************************
*****************************************************************************************************

Everything below this point is optional. It is NOT needed for an entry-level interview.
However, by studying these, you'll get greater exposure to more CS concepts and will be better prepared for
any software engineering job. You'll be a much more well-rounded software engineer.

*****************************************************************************************************
*****************************************************************************************************

⬆ back to top

Additional Books

These are here so you can dive into a topic you find interesting.
  • The Unix Programming Environment
    • An oldie but a goodie
  • The Linux Command Line: A Complete Introduction
    • A modern option
  • TCP/IP Illustrated Series
  • Head First Design Patterns
    • A gentle introduction to design patterns
  • Design Patterns: Elements of Reusable Object-Oriented Software
    • AKA the "Gang Of Four" book or GOF
    • The canonical design patterns book
  • Algorithm Design Manual (Skiena)
    • As a review and problem-recognition
    • The algorithm catalog portion is well beyond the scope of difficulty you'll get in an interview
    • This book has 2 parts:
      • Class textbook on data structures and algorithms
        • Pros:
          • Is a good review as any algorithms textbook would be
          • Nice stories from his experiences solving problems in industry and academia
          • Code examples in C
        • Cons:
          • Can be as dense or impenetrable as CLRS, and in some cases, CLRS may be a better alternative for some subjects
          • Chapters 7, 8, and 9 can be painful to try to follow, as some items are not explained well or require more brain than I have
          • Don't get me wrong: I like Skiena, his teaching style, and mannerisms, but I may not be Stony Brook material
      • Algorithm catalog:
        • This is the real reason you buy this book.
        • This book is better as an algorithm reference, and not something you read cover to cover.
    • Can rent it on Kindle
    • Answers:
    • Errata
  • Algorithm (Jeff Erickson)
  • Write Great Code: Volume 1: Understanding the Machine
    • The book was published in 2004, and is somewhat outdated, but it's a terrific resource for understanding a computer in brief
    • The author invented HLA, so take mentions and examples in HLA with a grain of salt. Not widely used, but decent examples of what assembly looks like
    • These chapters are worth the read to give you a nice foundation:
      • Chapter 2 - Numeric Representation
      • Chapter 3 - Binary Arithmetic and Bit Operations
      • Chapter 4 - Floating-Point Representation
      • Chapter 5 - Character Representation
      • Chapter 6 - Memory Organization and Access
      • Chapter 7 - Composite Data Types and Memory Objects
      • Chapter 9 - CPU Architecture
      • Chapter 10 - Instruction Set Architecture
      • Chapter 11 - Memory Architecture and Organization
  • Introduction to Algorithms
    • Important: Reading this book will only have limited value. This book is a great review of algorithms and data structures, but won't teach you how to write good code. You have to be able to code a decent solution efficiently
    • AKA CLR, sometimes CLRS, because Stein was late to the game
  • Computer Architecture, Sixth Edition: A Quantitative Approach
    • For a richer, more up-to-date (2017), but longer treatment

⬆ back to top

System Design, Scalability, Data Handling

You can expect system design questions if you have 4+ years of experience.

⬆ back to top

Additional Learning

I added them to help you become a well-rounded software engineer and to be aware of certain
technologies and algorithms, so you'll have a bigger toolbox.

⬆ back to top

Additional Detail on Some Subjects

I added these to reinforce some ideas already presented above, but didn't want to include them
above because it's just too much. It's easy to overdo it on a subject.
You want to get hired in this century, right?

⬆ back to top

Video Series

Sit back and enjoy.

⬆ back to top

Computer Science Courses

⬆ back to top

Algorithms implementation

⬆ back to top

Papers

⬆ back to top

LICENSE

CC-BY-SA-4.0

▼ Show full
Skills

My Codex Skills

GitHub Pages

Skills Public

A collection of reusable development skills for Apple platforms, GitHub workflows, refactoring, diff review swarms, bug investigation swarms, code review, React performance work, and skill curation.

Overview

This repository contains focused, self-contained skills that help with recurring engineering tasks such as generating App Store release notes, debugging iOS apps, improving SwiftUI and React code, packaging macOS apps, running multi-agent diff reviews and bug hunts, reviewing and simplifying code changes, orchestrating larger refactors, and auditing what new skills a project actually needs.

Install: place these skill folders under $CODEX_HOME/skills

Skills

This repo currently includes 16 skills:

Skill Folder Description
App Store Changelog app-store-changelog Creates user-facing App Store release notes from git history by collecting changes since the last tag, filtering for user-visible work, and rewriting it into concise "What's New" bullets.
GitHub github Uses the gh CLI to inspect and operate on GitHub issues, pull requests, workflow runs, and API data, including CI checks, run logs, and advanced queries.
iOS Debugger Agent ios-debugger-agent Uses XcodeBuildMCP to build, launch, and debug the current iOS app on a booted simulator, including UI inspection, interaction, screenshots, and log capture.
macOS Menubar Tuist App macos-menubar-tuist-app Builds, refactors, or reviews macOS menubar apps that use Tuist and SwiftUI, with emphasis on manifest ownership, store-layer architecture, and reliable local launch scripts.
macOS SwiftPM App Packaging (No Xcode) macos-spm-app-packaging Scaffolds, builds, packages, signs, and optionally notarizes SwiftPM-based macOS apps without requiring an Xcode project.
Orchestrate Batch Refactor orchestrate-batch-refactor Plans and executes larger refactor or rewrite efforts with dependency-aware parallel analysis and implementation using clearly scoped work packets.
Project Skill Audit project-skill-audit Analyzes a project's past Codex sessions, memory, existing local skills, and conventions to recommend the highest-value new skills or updates to existing ones.
React Component Performance react-component-performance Diagnoses slow React components by finding re-render churn, expensive render work, unstable props, and list bottlenecks, then suggests targeted optimizations and validation steps.
Bug Hunt Swarm bug-hunt-swarm Runs a read-only four-agent bug investigation focused on reproduction, code-path tracing, regressors, and the fastest proof step, then returns a ranked root-cause path.
Review and Simplify Changes review-and-simplify-changes Reviews a git diff or explicit file scope for reuse, code quality, efficiency, clarity, and standards issues, then optionally applies safe, behavior-preserving fixes.
Review Swarm review-swarm Runs a read-only four-agent diff review focused on behavioral regressions, security risks, performance or reliability issues, and contract or test coverage gaps, then returns a prioritized fix path.
Swift Concurrency Expert swift-concurrency-expert Reviews and fixes Swift 6.2+ concurrency issues such as actor isolation problems, Sendable violations, main-actor annotations, and data-race diagnostics.
SwiftUI Liquid Glass swiftui-liquid-glass Implements, reviews, or refactors SwiftUI features to use the iOS 26+ Liquid Glass APIs correctly, with proper modifier ordering, grouping, interactivity, and fallbacks.
SwiftUI Performance Audit swiftui-performance-audit Audits SwiftUI runtime performance from code and architecture, focusing on invalidation storms, identity churn, layout thrash, heavy render work, and profiling guidance.
SwiftUI UI Patterns swiftui-ui-patterns Provides best practices and example-driven guidance for building SwiftUI screens and components, including navigation, sheets, app wiring, async state, and reusable UI patterns.
SwiftUI View Refactor swiftui-view-refactor Refactors SwiftUI view files toward smaller subviews, MV-style data flow, stable view trees, explicit dependency injection, and correct Observation usage.

Usage

Each skill is self-contained. Refer to the SKILL.md file in each skill directory for triggers, workflow guidance, examples, and supporting references.

Contributing

Skills are designed to be focused and reusable. When adding new skills, ensure they:

  • Have a clear, single purpose
  • Include comprehensive documentation
  • Follow consistent patterns with existing skills
  • Include reference materials when applicable
▼ Show full
neovim

Vim-fork focused on extensibility and usability

Neovim

Documentation |
Chat

Coverity Scan analysis
Packages
Debian CI
Downloads

Neovim is a project that seeks to aggressively refactor Vim in order to:

See the Introduction wiki page and Roadmap
for more information.

Features

See :help nvim-features for the full list, and :help news for noteworthy changes in the latest version!

Install from package

Pre-built packages for Windows, macOS, and Linux are found on the
Releases page.

Managed packages are in Homebrew, Debian, Ubuntu, Fedora, Arch Linux, Void Linux, Gentoo, and more!

Install from source

See BUILD.md and supported platforms for details.

The build is CMake-based, but a Makefile is provided as a convenience.
After installing the dependencies, run the following command.

make CMAKE_BUILD_TYPE=RelWithDebInfo
sudo make install

To install to a non-default location:

make CMAKE_BUILD_TYPE=RelWithDebInfo CMAKE_INSTALL_PREFIX=/full/path/
make install

CMake hints for inspecting the build:

  • cmake --build build --target help lists all build targets.
  • build/CMakeCache.txt (or cmake -LAH build/) contains the resolved values of all CMake variables.
  • build/compile_commands.json shows the full compiler invocations for each translation unit.

Transitioning from Vim

See :help nvim-from-vim for instructions.

Project layout

├─ cmake/           CMake utils
├─ cmake.config/    CMake defines
├─ cmake.deps/      subproject to fetch and build dependencies (optional)
├─ runtime/         plugins and docs
├─ src/nvim/        application source code (see src/nvim/README.md)
│  ├─ api/          API subsystem
│  ├─ eval/         Vimscript subsystem
│  ├─ event/        event-loop subsystem
│  ├─ generators/   code generation (pre-compilation)
│  ├─ lib/          generic data structures
│  ├─ lua/          Lua subsystem
│  ├─ msgpack_rpc/  RPC subsystem
│  ├─ os/           low-level platform code
│  └─ tui/          built-in UI
└─ test/            tests (see test/README.md)

License

Neovim contributions since b17d96 are licensed under the
Apache 2.0 license, except for contributions copied from Vim (identified by the
vim-patch token). See LICENSE.txt for details.

▼ Show full
TaxHacker

Self-hosted AI accounting app. LLM analyzer for receipts, invoices, transactions with custom prompts and categories


TaxHacker — self-hosted AI accountant

GitHub Stars
License
GitHub Issues
Donate

🙏 I'm currently looking for a job! Here's my CV and my Github profile.

TaxHacker is a self-hosted accounting app designed for freelancers, indie hackers, and small businesses who want to save time and automate expense and income tracking using the power of modern AI.

Upload photos of receipts, invoices, or PDFs, and TaxHacker will automatically recognize and extract all the important data you need for accounting: product names, amounts, items, dates, merchants, taxes, and save it into a structured Excel-like database. You can even create custom fields with your own AI prompts to extract any specific information you need.

The app features automatic currency conversion (including crypto!) based on historical exchange rates from the transaction date. With built-in filtering, multi-project support, import/export capabilities, and custom categories, TaxHacker simplifies reporting and makes tax filing a bit easier.

🎥 Watch demo video

Dashboard

Important

This project is still in early development. Use at your own risk! Star us to get notified about new features and bugfixes ⭐️

✨ Features

1 Analyze photos and invoices with AI

Currency Conversion

Snap a photo of any receipt or upload an invoice PDF, and TaxHacker will automatically recognize, extract, categorize, and store all the information in a structured database.

  • Upload and organize your docs: Store multiple documents in "unsorted" until you're ready to process them manually or with AI assistance
  • AI data extraction: Use AI to automatically pull key information like dates, amounts, vendors, and line items
  • Auto-categorization: Transactions are automatically sorted into relevant categories based on their content
  • Item splitting: Extract individual items from invoices and split them into separate transactions when needed
  • Structured storage: Everything gets saved in an organized database for easy filtering and retrieval
  • Customizable AI providers: Choose from OpenAI, Google Gemini, or Mistral (local LLM support coming soon)

TaxHacker works with a wide variety of documents, including store receipts, restaurant bills, invoices, bank statements, letters, even handwritten receipts. It handles any language and any currency with ease.

2 Multi-currency support with automatic conversion (even crypto!)

Currency Conversion

TaxHacker automatically detects currencies in your documents and converts them to your base currency using historical exchange rates.

  • Foreight currency detection: Automatically identify the currency used in any document
  • Historical rates: Get conversion rates from the actual transaction date
  • All-world coverage: Support for 170+ world currencies and 14 popular cryptocurrencies (BTC, ETH, LTC, DOT, and more)
  • Flexible input: Manual entry is always available when you need more control

3 Organize your transactions using fully customizable categories, projects and fields

Transactions Table

Adapt TaxHacker to your unique needs with unlimited customization options. Create custom fields, projects, and categories that better suit your specific needs, idustry standards or country.

  • Custom categories and projecst: Create your own categories and projects to group your transactions in any convenient way
  • Custom fields: You can create unlimited number of custom fields to extraxt more information from your invoices (it's like creating extra columns in Excel)
  • Full-text search: Search through the actual content of recognized documents
  • Advanced filtering: Find exactly what you need with search and filter options
  • AI-powered extraction: Write your own prompts to extract any custom information from documents
  • Bulk operations: Process multiple documents or transactions at once

4 Customize any LLM prompt. Even system ones

Custom Categories

Take full control of how TaxHacker's AI processes your documents. Write custom AI prompts for fields, categories, and projects, or modify the built-in ones to match your specific needs.

  • Customizable system prompts: Modify the general prompt template in settings to suit your business
  • Field or project-specific prompts: Create custom extraction rules for your industry-specific documents
  • Full control: Adjust field extraction priorities and naming conventions to match your workflow
  • Industry optimization: Fine-tune the AI to understand your specific type of business documents
  • Full transparency: Every aspect of the AI extraction process is under your control and can be changed right in settings

TaxHacker is 100% adaptable and tunable to your unique requirements — whether you need to extract emails, addresses, project codes, or any other custom information from your documents.

5 Flexible data filtering and export

Data Export

Once your documents are processed, easily view, filter, and export your complete transaction history exactly how you need it.

  • Advanced filtering: Filter by date ranges, categories, projects, amounts, and any custom fields
  • Flexible exports: Export filtered transactions to CSV with all attached documents included
  • Tax-ready reports: Generate comprehensive reports for your accountant or tax advisor
  • Data portability: Download complete data archives to migrate to other services—your data stays yours

6 Self-hosted mode for data privacy

Self-hosting

Keep complete control over your financial data with local storage and self-hosting options. TaxHacker respects your privacy and gives you full ownership of your information.

  • Home server ready: Host on your own infrastructure for maximum privacy and control
  • Docker native: Simple setup with provided Docker containers and compose files
  • Data ownership: Your financial documents never leaves your control
  • No vendor lock-in: Export everything and migrate whenever you want
  • Transparent operations: Full access to source code and complete operational transparency

🛳 Deployment and Self-hosting

TaxHacker can be easily self-hosted on your own infrastructure for complete control over your data and application environment. We provide a Docker image and Docker Compose setup that makes deployment simple:

curl -O https://raw.githubusercontent.com/vas3k/TaxHacker/main/docker-compose.yml

docker compose up

The Docker Compose setup includes:

  • TaxHacker application container
  • PostgreSQL 17 database (or connect to your existing database)
  • Automatic database migrations on startup
  • Volume mounts for persistent data storage
  • Production-ready configuration

New Docker images are automatically built and published with every release. You can use specific version tags (e.g., v1.0.0) or latest for the most recent version.

For advanced setups, you can customize the Docker Compose configuration to fit your infrastructure. The default configuration uses the pre-built image from GitHub Container Registry, but you can also build locally using the provided Dockerfile.

Example custom configuration:

services:
  app:
    image: ghcr.io/vas3k/taxhacker:latest
    ports:
      - "7331:7331"
    environment:
      - SELF_HOSTED_MODE=true
      - UPLOAD_PATH=/app/data/uploads
      - DATABASE_URL=postgresql://postgres:postgres@localhost:5432/taxhacker
    volumes:
      - ./data:/app/data
    restart: unless-stopped

Environment Variables

Configure TaxHacker for your specific needs with these environment variables:

Variable Required Description Example
UPLOAD_PATH Yes Local directory for file uploads and storage ./data/uploads
DATABASE_URL Yes PostgreSQL connection string postgresql://user@localhost:5432/taxhacker
PORT No Port to run the application on 7331 (default)
BASE_URL No Base URL for the application http://localhost:7331
SELF_HOSTED_MODE No Set to "true" for self-hosting: enables auto-login, custom API keys, and additional features true
DISABLE_SIGNUP No Disable new user registration on your instance false
BETTER_AUTH_SECRET Yes Secret key for authentication (minimum 16 characters) your-secure-random-key

You can also configure LLM provider settings in the application or via environment variables:

  • OpenAI: OPENAI_MODEL_NAME and OPENAI_API_KEY
  • Google Gemini: GOOGLE_MODEL_NAME and GOOGLE_API_KEY
  • Mistral: MISTRAL_MODEL_NAME and MISTRAL_API_KEY

⌨️ Local Development

We use:

  • Next.js 15+ for the frontend and API
  • Prisma for database models and migrations
  • PostgreSQL as the database (PostgreSQL 17+ recommended)
  • Ghostscript and GraphicsMagick for PDF processing (install on macOS via brew install gs graphicsmagick)

Set up your local development environment:

# Clone the repository
git clone https://github.com/vas3k/TaxHacker.git
cd TaxHacker

# Install dependencies
npm install

# Set up environment variables
cp .env.example .env

# Edit .env with your configuration
# Make sure to set DATABASE_URL to your PostgreSQL connection string
# Example: postgresql://user@localhost:5432/taxhacker

# Initialize the database
npx prisma generate && npx prisma migrate dev

# Start the development server
npm run dev

Visit http://localhost:7331 to see your local TaxHacker instance in action.

For a production build, instead of npm run dev use the following commands:

# Build the application
npm run build

# Start the production server
npm run start

🤝 Contributing

We welcome contributions to TaxHacker! Here's how you can help make it even better:

  • 🐛 Bug Reports: File detailed issues when you encounter problems
  • 💡 Feature Requests: Share your ideas for new features and improvements
  • 🔧 Code Contributions: Submit pull requests to improve the application
  • 📚 Documentation: Help improve documentation and guides
  • 🎥 Content Creation: Videos, tutorials, and reviews help us reach more users!

All development happens on GitHub through issues and pull requests. We appreciate any help.

PRs Welcome

❤️ Support the Project

If TaxHacker has helped you save time or manage your finances better, consider supporting its continued development! Your donations help us maintain the project, add new features, and keep it free and open source. Every contribution helps ensure we can keep improving and maintaining this tool for the community.

Thank the TaxHacker devs

📄 License

TaxHacker is licensed under the MIT License.

▼ Show full
PaddleOCR

Turn any PDF or image document into structured data for your AI. A powerful, lightweight OCR toolkit that bridges the gap between images/PDFs and LLMs. Supports 100+ languages.

Star-history

Global Leading OCR Toolkit & Document AI Engine

English | 简体中文 | 繁體中文 | 日本語 | 한국어 | Français | Русский | Español | العربية

PyPI Downloads
Used by
python
os
hardware

AI Studio
Ask DeepWiki
License

PaddleOCR converts PDF documents and images into structured, LLM-ready data (JSON/Markdown) with industry-leading accuracy. With 70k+ Stars and trusted by top-tier projects like Dify, RAGFlow, and Cherry Studio, PaddleOCR is the bedrock for building intelligent RAG and Agentic applications.

🚀 Key Features

📄 Intelligent Document Parsing (LLM-Ready)

Transforming messy visuals into structured data for the LLM era.

  • SOTA Document VLM: Featuring PaddleOCR-VL-1.5 (0.9B), the industry's leading lightweight vision-language model for document parsing. It excels in parsing complex documents across 5 major "Real-World" challenges: Warping, Scanning, Screen Photography, Illumination, and Skewed documents, with structured outputs in Markdown and JSON formats.
  • Structure-Aware Conversion: Powered by PP-StructureV3, seamlessly convert complex PDFs and images into Markdown or JSON. Unlike the PaddleOCR-VL series models, it provides more fine-grained coordinate information, including table cell coordinates, text coordinates, and more.
  • Production-Ready Efficiency: Achieve commercial-grade accuracy with an ultra-small footprint. Outperforms numerous closed-source solutions in public benchmarks while remaining resource-efficient for edge/cloud deployment.

🔍 Universal Text Recognition (Scene OCR)

The global gold standard for high-speed, multilingual text spotting.

  • 100+ Languages Supported: Native recognition for a vast global library. Our PP-OCRv5 single-model solution elegantly handles multilingual mixed documents (Chinese, English, Japanese, Pinyin, etc.).
  • Complex Element Mastery: Beyond standard text recognition, we support natural scene text spotting across a wide range of environments, including IDs, street views, books, and industrial components
  • Performance Leap: PP-OCRv5 delivers a 13% accuracy boost over previous versions, maintaining the "Extreme Efficiency" that PaddleOCR is famous for.

PaddleOCR Architecture

🛠️ Developer-Centric Ecosystem

  • Seamless Integration: The premier choice for the AI Agent ecosystem—deeply integrated with Dify, RAGFlow, Pathway, and Cherry Studio.
  • LLM Data Flywheel: A complete pipeline to build high-quality datasets, providing a sustainable "Data Engine" for fine-tuning Large Language Models.
  • One-Click Deployment: Supports various hardware backends (NVIDIA GPU, Intel CPU, Kunlunxin XPU, and diverse AI Accelerators).

📣 Recent updates

🔥 [2026.01.29] PaddleOCR v3.4.0 Released: The Era of Irregular Document Parsing

  • PaddleOCR-VL-1.5 (SOTA 0.9B VLM): Our latest flagship model for document parsing is now live!
    • 94.5% Accuracy on OmniDocBench: Surpassing top-tier general large models and specialized document parsers.
    • Real-World Robustness: First to introduce the PP-DocLayoutV3 algorithm for irregular shape positioning, mastering 5 tough scenarios: Skew, Warping, Scanning, Illumination, and Screen Photography.
    • Capability Expansion: Now supports Seal Recognition, Text Spotting, and expands to 111 languages (including China’s Tibetan script and Bengali).
    • Long Document Mastery: Supports automatic cross-page table merging and hierarchical heading identification.
    • Try it now: Available on HuggingFace or our Official Website.
2025.10.16: Release of PaddleOCR 3.3.0

  • Released PaddleOCR-VL:

    • Model Introduction:

      • PaddleOCR-VL is a SOTA and resource-efficient model tailored for document parsing. Its core component is PaddleOCR-VL-0.9B, a compact yet powerful vision-language model (VLM) that integrates a NaViT-style dynamic resolution visual encoder with the ERNIE-4.5-0.3B language model to enable accurate element recognition. This innovative model efficiently supports 109 languages and excels in recognizing complex elements (e.g., text, tables, formulas, and charts), while maintaining minimal resource consumption. Through comprehensive evaluations on widely used public benchmarks and in-house benchmarks, PaddleOCR-VL achieves SOTA performance in both page-level document parsing and element-level recognition. It significantly outperforms existing solutions, exhibits strong competitiveness against top-tier VLMs, and delivers fast inference speeds. These strengths make it highly suitable for practical deployment in real-world scenarios. The model has been released on HuggingFace. Everyone is welcome to download and use it! More introduction infomation can be found in PaddleOCR-VL.

    • Core Features:

      • Compact yet Powerful VLM Architecture: We present a novel vision-language model that is specifically designed for resource-efficient inference, achieving outstanding performance in element recognition. By integrating a NaViT-style dynamic high-resolution visual encoder with the lightweight ERNIE-4.5-0.3B language model, we significantly enhance the model’s recognition capabilities and decoding efficiency. This integration maintains high accuracy while reducing computational demands, making it well-suited for efficient and practical document processing applications.
      • SOTA Performance on Document Parsing: PaddleOCR-VL achieves state-of-the-art performance in both page-level document parsing and element-level recognition. It significantly outperforms existing pipeline-based solutions and exhibiting strong competitiveness against leading vision-language models (VLMs) in document parsing. Moreover, it excels in recognizing complex document elements, such as text, tables, formulas, and charts, making it suitable for a wide range of challenging content types, including handwritten text and historical documents. This makes it highly versatile and suitable for a wide range of document types and scenarios.
      • Multilingual Support: PaddleOCR-VL Supports 109 languages, covering major global languages, including but not limited to Chinese, English, Japanese, Latin, and Korean, as well as languages with different scripts and structures, such as Russian (Cyrillic script), Arabic, Hindi (Devanagari script), and Thai. This broad language coverage substantially enhances the applicability of our system to multilingual and globalized document processing scenarios.

  • Released PP-OCRv5 Multilingual Recognition Model:

    • Improved the accuracy and coverage of Latin script recognition; added support for Cyrillic, Arabic, Devanagari, Telugu, Tamil, and other language systems, covering recognition of 109 languages. The model has only 2M parameters, and the accuracy of some models has increased by over 40% compared to the previous generation.
2025.08.21: Release of PaddleOCR 3.2.0

  • Significant Model Additions:

    • Introduced training, inference, and deployment for PP-OCRv5 recognition models in English, Thai, and Greek. The PP-OCRv5 English model delivers an 11% improvement in English scenarios compared to the main PP-OCRv5 model, with the Thai and Greek recognition models achieving accuracies of 82.68% and 89.28%, respectively.

  • Deployment Capability Upgrades:

    • Full support for PaddlePaddle framework versions 3.1.0 and 3.1.1.
    • Comprehensive upgrade of the PP-OCRv5 C++ local deployment solution, now supporting both Linux and Windows, with feature parity and identical accuracy to the Python implementation.
    • High-performance inference now supports CUDA 12, and inference can be performed using either the Paddle Inference or ONNX Runtime backends.
    • The high-stability service-oriented deployment solution is now fully open-sourced, allowing users to customize Docker images and SDKs as required.
    • The high-stability service-oriented deployment solution also supports invocation via manually constructed HTTP requests, enabling client-side code development in any programming language.

  • Benchmark Support:

    • All production lines now support fine-grained benchmarking, enabling measurement of end-to-end inference time as well as per-layer and per-module latency data to assist with performance analysis. Here's how to set up and use the benchmark feature.
    • Documentation has been updated to include key metrics for commonly used configurations on mainstream hardware, such as inference latency and memory usage, providing deployment references for users.

  • Bug Fixes:

    • Resolved the issue of failed log saving during model training.
    • Upgraded the data augmentation component for formula models for compatibility with newer versions of the albumentations dependency, and fixed deadlock warnings when using the tokenizers package in multi-process scenarios.
    • Fixed inconsistencies in switch behaviors (e.g., use_chart_parsing) in the PP-StructureV3 configuration files compared to other pipelines.

  • Other Enhancements:

    • Separated core and optional dependencies. Only minimal core dependencies are required for basic text recognition; additional dependencies for document parsing and information extraction can be installed as needed.
    • Enabled support for NVIDIA RTX 50 series graphics cards on Windows; users can refer to the installation guide for the corresponding PaddlePaddle framework versions.
    • PP-OCR series models now support returning single-character coordinates.
    • Added AIStudio, ModelScope, and other model download sources, allowing users to specify the source for model downloads.
    • Added support for chart-to-table conversion via the PP-Chart2Table module.
    • Optimized documentation descriptions to improve usability.

History Log

🚀 Quick Start

Step 1: Try Online

PaddleOCR official website provides interactive Experience Center and APIs—no setup required, just one click to experience.

👉 Visit Official Website

Step 2: Local Deployment

For local usage, please refer to the following documentation based on your needs:

🧩 More Features

🔄 Quick Overview of Execution Results

PP-OCRv5

PP-OCRv5 Demo

PP-StructureV3

PP-StructureV3 Demo

PaddleOCR-VL

PP-StructureV3 Demo

✨ Stay Tuned

⭐ Star this repository to keep up with exciting updates and new releases, including powerful OCR and document parsing capabilities! ⭐

Star-Project

👩‍👩‍👧‍👦 Community

PaddlePaddle WeChat official account Join the tech discussion group

😃 Awesome Projects Leveraging PaddleOCR

PaddleOCR wouldn't be where it is today without its incredible community! 💗 A massive thank you to all our longtime partners, new collaborators, and everyone who's poured their passion into PaddleOCR — whether we've named you or not. Your support fuels our fire!

Project Name Description
Dify Production-ready platform for agentic workflow development.
RAGFlow RAG engine based on deep document understanding.
pathway Python ETL framework for stream processing, real-time analytics, LLM pipelines, and RAG.
MinerU Multi-type Document to Markdown Conversion Tool
Umi-OCR Free, Open-source, Batch Offline OCR Software.
cherry-studio A desktop client that supports for multiple LLM providers.
haystack AI orchestration framework to build customizable, production-ready LLM applications.
OmniParser OmniParser: Screen Parsing tool for Pure Vision Based GUI Agent.
QAnything Question and Answer based on Anything.
Learn more projects More projects based on PaddleOCR

👩‍👩‍👧‍👦 Contributors

🌟 Star

Star-history

📄 License

This project is released under the Apache 2.0 license.

🎓 Citation

@misc{cui2025paddleocr30technicalreport,
      title={PaddleOCR 3.0 Technical Report}, 
      author={Cheng Cui and Ting Sun and Manhui Lin and Tingquan Gao and Yubo Zhang and Jiaxuan Liu and Xueqing Wang and Zelun Zhang and Changda Zhou and Hongen Liu and Yue Zhang and Wenyu Lv and Kui Huang and Yichao Zhang and Jing Zhang and Jun Zhang and Yi Liu and Dianhai Yu and Yanjun Ma},
      year={2025},
      eprint={2507.05595},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2507.05595}, 
}

@misc{cui2025paddleocrvlboostingmultilingualdocument,
      title={PaddleOCR-VL: Boosting Multilingual Document Parsing via a 0.9B Ultra-Compact Vision-Language Model}, 
      author={Cheng Cui and Ting Sun and Suyin Liang and Tingquan Gao and Zelun Zhang and Jiaxuan Liu and Xueqing Wang and Changda Zhou and Hongen Liu and Manhui Lin and Yue Zhang and Yubo Zhang and Handong Zheng and Jing Zhang and Jun Zhang and Yi Liu and Dianhai Yu and Yanjun Ma},
      year={2025},
      eprint={2510.14528},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2510.14528}, 
}

@misc{cui2026paddleocrvl15multitask09bvlm,
      title={PaddleOCR-VL-1.5: Towards a Multi-Task 0.9B VLM for Robust In-the-Wild Document Parsing}, 
      author={Cheng Cui and Ting Sun and Suyin Liang and Tingquan Gao and Zelun Zhang and Jiaxuan Liu and Xueqing Wang and Changda Zhou and Hongen Liu and Manhui Lin and Yue Zhang and Yubo Zhang and Yi Liu and Dianhai Yu and Yanjun Ma},
      year={2026},
      eprint={2601.21957},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2601.21957}, 
}
▼ Show full
ChatDev

ChatDev 2.0: Dev All through LLM-powered Multi-Agent Collaboration

ChatDev 2.0 - DevAll

DevAll Logo

A Zero-Code Multi-Agent Platform for Developing Everything

【English | 简体中文】

【📚 Developers | 👥 Contributors｜⭐️ ChatDev 1.0 (Legacy)】

📖 Overview

ChatDev has evolved from a specialized software development multi-agent system into a comprehensive multi-agent orchestration platform.

  • ChatDev 2.0 (DevAll) is a Zero-Code Multi-Agent Platform for "Developing Everything". It empowers users to rapidly build and execute customized multi-agent systems through simple configuration. No coding is required—users can define agents, workflows, and tasks to orchestrate complex scenarios such as data visualization, 3D generation, and deep research.
  • ChatDev 1.0 (Legacy) operates as a Virtual Software Company. It utilizes various intelligent agents (e.g., CEO, CTO, Programmer) participating in specialized functional seminars to automate the entire software development life cycle—including designing, coding, testing, and documenting. It serves as the foundational paradigm for communicative agent collaboration.

🎉 News

• Jan 07, 2026: 🚀 We are excited to announce the official release of ChatDev 2.0 (DevAll)! This version introduces a zero-code multi-agent orchestration platform. The classic ChatDev (v1.x) has been moved to the chatdev1.0 branch for maintenance. More details about ChatDev 2.0 can be found on our official post.

Old News

•Sep 24, 2025: 🎉 Our paper Multi-Agent Collaboration via Evolving Orchestration has been accepted to NeurIPS 2025. The implementation is available in the puppeteer branch of this repository.

•May 26, 2025: 🎉 We propose a novel puppeteer-style paradigm for multi-agent collaboration among large language model based agents. By leveraging a learnable central orchestrator optimized with reinforcement learning, our method dynamically activates and sequences agents to construct efficient, context-aware reasoning paths. This approach not only improves reasoning quality but also reduces computational costs, enabling scalable and adaptable multi-agent cooperation in complex tasks.
See our paper in Multi-Agent Collaboration via Evolving Orchestration.

•June 25, 2024: 🎉To foster development in LLM-powered multi-agent collaboration🤖🤖 and related fields, the ChatDev team has curated a collection of seminal papers📄 presented in a open-source interactive e-book📚 format. Now you can explore the latest advancements on the Ebook Website and download the paper list.

•June 12, 2024: We introduced Multi-Agent Collaboration Networks (MacNet) 🎉, which utilize directed acyclic graphs to facilitate effective task-oriented collaboration among agents through linguistic interactions 🤖🤖. MacNet supports co-operation across various topologies and among more than a thousand agents without exceeding context limits. More versatile and scalable, MacNet can be considered as a more advanced version of ChatDev's chain-shaped topology. Our preprint paper is available at https://arxiv.org/abs/2406.07155. This technique has been incorporated into the macnet branch, enhancing support for diverse organizational structures and offering richer solutions beyond software development (e.g., logical reasoning, data analysis, story generation, and more).

• May 07, 2024, we introduced "Iterative Experience Refinement" (IER), a novel method where instructor and assistant agents enhance shortcut-oriented experiences to efficiently adapt to new tasks. This approach encompasses experience acquisition, utilization, propagation and elimination across a series of tasks and making the pricess shorter and efficient. Our preprint paper is available at https://arxiv.org/abs/2405.04219, and this technique will soon be incorporated into ChatDev.

• January 25, 2024: We have integrated Experiential Co-Learning Module into ChatDev. Please see the Experiential Co-Learning Guide.

• December 28, 2023: We present Experiential Co-Learning, an innovative approach where instructor and assistant agents accumulate shortcut-oriented experiences to effectively solve new tasks, reducing repetitive errors and enhancing efficiency. Check out our preprint paper at https://arxiv.org/abs/2312.17025 and this technique will soon be integrated into ChatDev.

• November 15, 2023: We launched ChatDev as a SaaS platform that enables software developers and innovative entrepreneurs to build software efficiently at a very low cost and remove the barrier to entry. Try it out at https://chatdev.modelbest.cn/.

• November 2, 2023: ChatDev is now supported with a new feature: incremental development, which allows agents to develop upon existing codes. Try --config "incremental" --path "[source_code_directory_path]" to start it.

• October 26, 2023: ChatDev is now supported with Docker for safe execution (thanks to contribution from ManindraDeMel). Please see Docker Start Guide.

• September 25, 2023: The Git mode is now available, enabling the programmer to utilize Git for version control. To enable this feature, simply set "git_management" to "True" in ChatChainConfig.json. See guide.

• September 20, 2023: The Human-Agent-Interaction mode is now available! You can get involved with the ChatDev team by playing the role of reviewer and making suggestions to the programmer ;
try python3 run.py --task [description_of_your_idea] --config "Human". See guide and example.

• September 1, 2023: The Art mode is available now! You can activate the designer agent to generate images used in the software;
try python3 run.py --task [description_of_your_idea] --config "Art". See guide and example.

• August 28, 2023: The system is publicly available.

• August 17, 2023: The v1.0.0 version was ready for release.

• July 30, 2023: Users can customize ChatChain, Phasea and Role settings. Additionally, both online Log mode and replay
mode are now supported.

• July 16, 2023: The preprint paper associated with this project was published.

• June 30, 2023: The initial version of the ChatDev repository was released.

🚀 Quick Start

📋 Prerequisites

  • OS: macOS / Linux / WSL / Windows
  • Python: 3.12+
  • Node.js: 18+
  • Package Manager: uv

📦 Installation

  1. Backend Dependencies (Python managed by uv):

    uv sync

  2. Frontend Dependencies (Vite + Vue 3):

    cd frontend && npm install

🔑 Configuration

  • Environment Variables:
    cp .env.example .env
  • Model Keys: Set API_KEY and BASE_URL in .env for your LLM provider.
  • YAML placeholders: Use ${VAR}（e.g., ${API_KEY}）in configuration files to reference these variables.

⚡️ Run the Application

Using Makefile (Recommended)

Start both Backend and Frontent:

make dev

Then access the Web Console at http://localhost:5173.

Manual Commands

  1. Start Backend:

    # Run from the project root
uv run python server_main.py --port 6400 --reload

    Remove --reload if output files (e.g., GameDev) trigger restarts, which interrupts tasks and loses progress.

  2. Start Frontend:

    cd frontend
VITE_API_BASE_URL=http://localhost:6400 npm run dev

    Then access the Web Console at http://localhost:5173.

    💡 Tip: If the frontend fails to connect to the backend, the default port 6400 may already be occupied.
    Please switch both services to an available port, for example:

    • Backend: start with --port 6401
    • Frontend: set VITE_API_BASE_URL=http://localhost:6401

Utility Commands

  • Help command:

    make help

  • Sync YAML workflows to frontend:

    make sync

    Uploads all workflow files from yaml_instance/ to the database.

  • Validate all YAML workflows:

    make validate-yamls

    Checks all YAML files for syntax and schema errors.

🦞 Run with OpenClaw

OpenClaw can integrate with ChatDev by invoking existing agent teams or dynamically creating new agent teams within ChatDev.
To get started:

  1. Start the ChatDev 2.0 backend.

  2. Install the required skills for your OpenClaw instance:

    clawdhub install chatdev

  3. Ask your OpenClaw to create a ChatDev workflow. For example:

  • Automated information collection and content publishing

    Create a ChatDev workflow to automatically collect trending information, generate a Xiaohongshu post, and publish it.

  • Multi-agent geopolitical simulation

    Create a ChatDev workflow with multiple agents to simulate possible future developments of the Middle East situation.

🐳 Run with Docker

Alternatively, you can run the entire application using Docker Compose. This method simplifies dependency management and provides a consistent environment.

  1. Prerequisites:

    • Docker and Docker Compose installed.
    • Ensure you have a .env file in the project root for your API keys.

  2. Build and Run:

    # From the project root
docker compose up --build

  3. Access:

    • Backend: http://localhost:6400
    • Frontend: http://localhost:5173

The services will automatically restart if they crash, and local file changes will be reflected inside the containers for live development.

💡 How to Use

🖥️ Web Console

The DevAll interface provides a seamless experience for both construction and execution

  • Tutorial: Comprehensive step-by-step guides and documentation integrated directly into the platform to help you get started quickly.

  • Workflow: A visual canvas to design your multi-agent systems. Configure node parameters, define context flows, and orchestrate complex agent interactions with drag-and-drop ease.

  • Launch: Initiate workflows, monitor real-time logs, inspect intermediate artifacts, and provide human-in-the-loop feedback.

🧰 Python SDK

For automation and batch processing, use our lightweight Python SDK to execute workflows programmatically and retrieve results directly.

from runtime.sdk import run_workflow

# Execute a workflow and get the final node message
result = run_workflow(
    yaml_file="yaml_instance/demo.yaml",
    task_prompt="Summarize the attached document in one sentence.",
    attachments=["/path/to/document.pdf"],
    variables={"API_KEY": "sk-xxxx"} # Override .env variables if needed
)

if result.final_message:
    print(f"Output: {result.final_message.text_content()}")

We have released the ChatDev Python SDK (PyPI package chatdev), so you can also run YAML workflow and multi-agent tasks directly in Python. For installation and version details, see PyPI: chatdev 0.1.0.

⚙️ For Developers

For secondary development and extensions, please proceed with this section.

Extend DevAll with new nodes, providers, and tools.
The project is organized into a modular structure:

  • Core Systems: server/ hosts the FastAPI backend, while runtime/ manages agent abstraction and tool execution.
  • Orchestration: workflow/ handles the multi-agent logic, driven by configurations in entity/.
  • Frontend: frontend/ contains the Vue 3 Web Console.
  • Extensibility: functions/ is the place for custom Python tools.

Relevant reference documentation:

🌟 Featured Workflows

We provide robust, out-of-the-box templates for common scenarios. All runnable workflow configs are located in yaml_instance/.

  • Demos: Files named demo_*.yaml showcase specific features or modules.
  • Implementations: Files named directly (e.g., ChatDev_v1.yaml) are full in-house or recreated workflows. As follows:

📋 Workflow Collection

Category Workflow Case
📈 Data Visualization data_visualization_basic.yaml
data_visualization_enhanced.yaml
Prompt: "Create 4–6 high-quality PNG charts for my large real-estate transactions dataset."
🛠️ 3D Generation
(Requires Blender & blender-mcp)		 blender_3d_builder_simple.yaml
blender_3d_builder_hub.yaml
blender_scientific_illustration.yaml
Prompt: "Please build a Christmas tree."
🎮 Game Dev GameDev_v1.yaml
ChatDev_v1.yaml
Prompt: "Please help me design and develop a Tank Battle game."
📚 Deep Research deep_research_v1.yaml
Prompt: "Research about recent advances in the field of LLM-based agent RL"
🎓 Teach Video teach_video.yaml (Please run command uv add manim before running this workflow)
Prompt: "讲一下什么是凸优化"

💡 Usage Guide

For those implementations, you can use the Launch tab to execute them.

  1. Select: Choose a workflow in the Launch tab.
  2. Upload: Upload necessary files (e.g., .csv for data analysis) if required.
  3. Prompt: Enter your request (e.g., "Visualize the sales trends" or "Design a snake game").

🤝 Contributing

We welcome contributions from the community! Whether you're fixing bugs, adding new workflow templates, or sharing high-quality cases/artifacts produced by DevAll, your help is much appreciated. Feel free to contribute by submitting Issues or Pull Requests.

By contributing to DevAll, you'll be recognized in our Contributors list below. Check out our Developer Guide to get started!

👥 Contributors

Primary Contributors


NA-Wen
zxrys
swugi
huatl98

Contributors


LaansDole
zivkovicp
shiowen
kilo2127
AckerlyLau
rainoeelmae
conprour
Br1an67
NINE-J
Yanghuabei

🤝 Acknowledgments

  
  
  



🔎 Citation

@article{chatdev,
    title = {ChatDev: Communicative Agents for Software Development},
    author = {Chen Qian and Wei Liu and Hongzhang Liu and Nuo Chen and Yufan Dang and Jiahao Li and Cheng Yang and Weize Chen and Yusheng Su and Xin Cong and Juyuan Xu and Dahai Li and Zhiyuan Liu and Maosong Sun},
    journal = {arXiv preprint arXiv:2307.07924},
    url = {https://arxiv.org/abs/2307.07924},
    year = {2023}
}

@article{colearning,
    title = {Experiential Co-Learning of Software-Developing Agents},
    author = {Chen Qian and Yufan Dang and Jiahao Li and Wei Liu and Zihao Xie and Yifei Wang and Weize Chen and Cheng Yang and Xin Cong and Xiaoyin Che and Zhiyuan Liu and Maosong Sun},
    journal = {arXiv preprint arXiv:2312.17025},
    url = {https://arxiv.org/abs/2312.17025},
    year = {2023}
}

@article{macnet,
    title={Scaling Large-Language-Model-based Multi-Agent Collaboration},
    author={Chen Qian and Zihao Xie and Yifei Wang and Wei Liu and Yufan Dang and Zhuoyun Du and Weize Chen and Cheng Yang and Zhiyuan Liu and Maosong Sun}
    journal={arXiv preprint arXiv:2406.07155},
    url = {https://arxiv.org/abs/2406.07155},
    year={2024}
}

@article{iagents,
    title={Autonomous Agents for Collaborative Task under Information Asymmetry},
    author={Wei Liu and Chenxi Wang and Yifei Wang and Zihao Xie and Rennai Qiu and Yufan Dnag and Zhuoyun Du and Weize Chen and Cheng Yang and Chen Qian},
    journal={arXiv preprint arXiv:2406.14928},
    url = {https://arxiv.org/abs/2406.14928},
    year={2024}
}

@article{puppeteer,
      title={Multi-Agent Collaboration via Evolving Orchestration}, 
      author={Yufan Dang and Chen Qian and Xueheng Luo and Jingru Fan and Zihao Xie and Ruijie Shi and Weize Chen and Cheng Yang and Xiaoyin Che and Ye Tian and Xuantang Xiong and Lei Han and Zhiyuan Liu and Maosong Sun},
      journal={arXiv preprint arXiv:2505.19591},
      url={https://arxiv.org/abs/2505.19591},
      year={2025}
}

📬 Contact

If you have any questions, feedback, or would like to get in touch, please feel free to reach out to us via email at qianc62@gmail.com

▼ Show full
huaren wxcity hn newmitbbs bedtime bili
new
jiandan
new
hdeal github
Prev day Apr 01 Wed Next day Today
DISCLAIMER: This information is provided "as is". The post content belongs to the original source.