Papers with Code Open-Source Revival: A New Hub for AI Research Under the Apache 2.0 License

Open-Source Revival: From Meta Shutdown to Community Resurrection

The Papers with Code website, once acquired and shut down by Meta, has now been officially open-sourced under the Apache 2.0 license, marking a community-driven revival. Developers, in collaboration with AI assistants like Claude and Codex, have spent the past few months building a new site that tracks trending AI research, benchmarks, and SOTA (State-of-the-Art) results. The project's goal is clear: to build the best paper interaction interface together with the community.

The open-source code is hosted in Hugging Face's GitHub repository (huggingface/paperswithcode), managed as a monorepo. This move means any developer can freely view, modify, and contribute to the code, continuing the open spirit of the original Papers with Code.

Papers with Code is an academic resource platform created in 2018 by Robert Stojnic and others. It was initially designed to link machine learning papers on arXiv with their corresponding code implementations, datasets, and leaderboards, allowing researchers to directly compare the performance of different methods on standard benchmarks. In 2021, Meta AI acquired it and brought it in-house. Although the site remained accessible afterward, updates gradually slowed and eventually stopped around 2024, prompting widespread regret in the community. That is why this open-source revival is seen as a "community takeover": while the original data assets and brand belong to Meta, the new project rebuilds core functionality in an independent codebase, continuing the original mission of open scientific infrastructure.

The Apache 2.0 license is a permissive open-source license that allows users to freely use, modify, and distribute code, including for commercial purposes, while requiring retention of copyright and license notices and including a patent grant clause. Unlike copyleft licenses such as GPL, Apache 2.0 does not require derivative works to be open-sourced, making it more friendly to enterprise participation and commercial integration. It is often used for infrastructure projects intended for widespread adoption.

Technical Architecture: A Simple and Efficient Modern Stack

According to the project description, the website's tech stack is quite straightforward: Python FastAPI for the backend and React for the frontend. The entire application is hosted on a single Hetzner VPS, with the code stored on the same server. This lightweight deployment approach reduces maintenance costs and makes it easier for contributors to understand the overall structure.

The frontend primarily visualizes paper information by calling various API endpoints, and these APIs are themselves publicly available. In addition, the project provides a companion MCP server, making it easy for other AI tools or agents to access paper data. For developers looking to integrate AI research tracking capabilities, this undoubtedly lowers the barrier to entry.

MCP (Model Context Protocol) is an open protocol proposed by Anthropic in late 2024 to enable AI models or agents to connect to external data sources and tools in a standardized way. Think of an MCP server as an "adapter": it wraps resources such as paper metadata and benchmark results into interfaces that models can call, so that MCP-compatible clients like Claude and Cursor do not need to write separate integration code for each data source. For this project, providing a companion MCP server means that any AI assistant compatible with the protocol can directly query Papers with Code's paper and SOTA data, rather than being limited to browser access.

FastAPI is a popular asynchronous web framework in the Python ecosystem in recent years. It automatically generates request validation and OpenAPI documentation based on type annotations, making it suitable for rapidly building RESTful APIs. React is a frontend component library maintained by Meta, known for its declarative UI and virtual DOM. Together they form a typical "frontend-backend separation" architecture: the backend is only responsible for data interfaces, and the frontend handles rendering. This also explains why the project can conveniently expose its APIs to third parties.

Community-Driven: From Personal Project to Open Collaboration

The developer mentioned that they had recently received many feature requests, and rather than developing alone, it would be better to make the codebase public so anyone can contribute. This approach aligns with the usual practices of the open-source community and echoes Papers with Code's original positioning as a community resource.

The original Papers with Code gradually ceased service after being acquired by Meta, which many researchers lamented. Today's new project not only replicates the core features—trending papers, benchmarks, and SOTA leaderboards—but also accelerates development with modern AI tools. The involvement of Claude and Codex also demonstrates the real-world application of AI-assisted programming in actual projects.

Monorepo refers to managing multiple subprojects—such as frontend, backend, and shared libraries—in a single Git repository rather than splitting them into separate repositories. This approach facilitates unified versioning, shared configuration, and one-time commits of cross-cutting changes. Many large open-source projects on GitHub adopt this model; the trade-off is a larger repository size, requiring tools for incremental builds and granular permissions. For contributors who want to improve both the API and the interface simultaneously, a monorepo significantly reduces collaboration friction.

Significance for the AI Research Ecosystem

The rebirth of Papers with Code is not just a technical revival but also represents a community-oriented return of research infrastructure. In an era of rapid AI research development, an open and customizable paper tracking platform can help researchers discover cutting-edge work faster and compare benchmark results. The public API and MCP server provide data entry points for automation tools and agents, potentially spawning more innovative applications.

However, the project is still in its early stages, and its feature completeness and data coverage remain to be tested by the community. The choice of open-source license (Apache 2.0) allows commercial use and modification, which is beneficial for ecosystem building, but the community also needs to jointly maintain code quality and data accuracy.

How to Participate

Interested developers can visit the GitHub repository to view the code, submit issues, or pull requests. Since the project uses a monorepo structure, contributing to either the frontend or backend is relatively convenient. The companion MCP server also provides ready-made interfaces for developers who want to build AI agents.

The open-source revival of Papers with Code may become a model for the community-driven development of AI research tools. In an era where large tech companies dominate infrastructure, this community-driven, AI-assisted development model is worth watching.

One-line Summary

Papers with Code, shut down by Meta, is revived as open source under Apache 2.0, with the community and AI collaborating to rebuild the paper tracking platform.

Paragraph Summary

Papers with Code, once acquired by Meta and discontinued, is now officially open-sourced under the Apache 2.0 license. The code is hosted in Hugging Face's GitHub repository using a monorepo structure. The new website is being built by developers in collaboration with AI assistants like Claude and Codex, aiming to track trending AI research, benchmarks, and SOTA leaderboards and to create the best paper interaction interface together with the community. The tech stack is a Python FastAPI backend plus a React frontend, deployed on a single Hetzner VPS, and it provides a public API and a companion MCP server to facilitate access to paper data by other AI tools. The article views this as a community-oriented return of research infrastructure, but the project is still early-stage, with features and data coverage yet to be validated.