# Open Paper: A Self-Hostable Research Workbench With Citation-Grounded AI > Open Paper is an open-source workspace for reading and annotating research papers, with an AI assistant that cites every answer by linking back to the exact passage in the source document. It is licensed under AGPL-3.0 and can be self-hosted, though the project's own documentation warns that self-hosting requires significant assembly. **khoj-ai/openpaper** — Open Paper is a workbench for managing your research library. Read, annotate, and understand your papers all in one place. Use an AI assistant to conduct your literature review. - Repository: https://github.com/khoj-ai/openpaper - Website: https://openpaper.ai - Stars: 483 · Forks: 72 - Language: Python - License: AGPL-3.0 - Published: 2026-09-15 · Updated: 2026-09-15 · Language: en - Canonical page: https://hysenlabs.com/projects/khoj-ai-openpaper ## The Problem Open Paper Addresses Reading research papers involves constant context switching: looking up unfamiliar terms, chasing cited work, and relocating a highlighted result from three papers ago. Existing tools split those tasks across separate applications. Open Paper puts the document viewer, annotation layer, AI assistant, and cross-paper search into one interface. The target users are researchers and graduate students who manage a library of PDFs and need to run literature reviews. The core value proposition is an AI assistant that sits beside the document rather than in a separate chat window, so the reader never leaves the paper. Every AI-generated answer includes citations that link directly to the passage in the paper it came from. The README describes the design principle as: trust, but verify. The project publishes a hosted version at openpaper.ai, so users who do not want to self-host can try the full feature set without installing anything. The open-source repository is the codebase behind that service. ## How the AI Assistant and Citation System Work When a PDF is uploaded, Open Paper immediately generates an AI summary and a set of starter questions, which the README describes as a way to ground the reader before diving into the document. As the reader works through the paper, the AI assistant is always visible in a side panel. Every response from the assistant uses contextual citations. Each citation is a clickable reference that jumps the document viewer to the exact passage the answer is drawn from. This differs from typical RAG setups where sources are named but not navigable. The team built a benchmark called ResearchQA, published at arxiv.org/abs/2607.11074, to measure how reliably the system cites accurately. The README links to an evaluation suite in server/evals/README.md for the methods, metrics, and how to reproduce the evaluation. For multi-paper work, papers can be grouped into projects. A project enables cross-paper queries and data tables: the user defines custom schemas, and Open Paper fills each cell from across every paper in the collection, with citations throughout. The data table output can be exported to CSV. ## Annotating Papers and Building a Library Highlighting any text in the document creates an annotation. The user can attach a comment to the highlight, send the selection to the AI assistant for a deeper explanation, or keep it as a note. Notes are displayed alongside the document with a markdown view available. Papers can be imported from Zotero by connecting a Zotero account; the integration syncs automatically after the initial connection. Papers can also be found through a built-in open-access literature search that lets the user pull results directly into their library with one click. Audio overviews can be generated for any paper for listening away from a screen. The README describes this feature as an option for when reading is not possible, without specifying which TTS backend is used. The underlying configuration is handled by the Celery worker described in jobs/README.md. ## Self-Hosting: What Is Required Open Paper is open source and self-hostable. The README is direct about the overhead: it is built primarily as a hosted service and is not optimised for self-hosting. The required components are PostgreSQL, S3-compatible object storage, LLM API keys, and background workers. The repository is structured as three services: - server/ contains the FastAPI backend. - client/ contains the Next.js web application. - jobs/ contains the Celery worker that handles PDF parsing, Zotero synchronisation, and audio generation. Each service has its own README. DEVELOPMENT.md in the repository root is the canonical starting point for getting a local environment running; it covers prerequisites, environment variables, and how to start all three services. The main README explicitly points to DEVELOPMENT.md for setup rather than including steps inline. The project has a docker-compose.dev.yml file alongside a docker/ directory for containerised development. The repository does not include a single-command production deployment path. ## License Implications of AGPL-3.0 Open Paper is licensed under the GNU Affero General Public License version 3. AGPL-3.0 extends GPL-style copyleft to network services: any organisation that runs a modified version of Open Paper over a network and lets users interact with it is required to make the modified source code available to those users. This is a meaningful constraint for teams considering Open Paper as the core of an internal or commercial research tool. For individual self-hosted use where no external users connect to the service, the practical effect is the same as GPL: modifications do not need to be published. For any hosted or shared deployment, they do. The hosted service at openpaper.ai operates under its own terms; the AGPL applies to the open-source codebase in this repository. ## Limitations and Cases Where Open Paper Is the Wrong Tool The self-hosting path requires significant infrastructure. A solo researcher who wants a local installation must set up and maintain Postgres, object storage, an LLM API subscription, and a Celery queue. For someone who only reads papers occasionally, the overhead is hard to justify when the hosted version at openpaper.ai covers the same feature set. The README does not document the precise LLM providers supported beyond what the self-hosting note implies (LLM API keys). Teams that want a fully offline deployment, with no external LLM calls, will need to audit the backend code rather than rely on documented configuration. A direct alternative for annotation-focused workflows is Zotero, which Open Paper can import from. Zotero handles citation management and PDF annotation but has no built-in AI assistant. The difference in approach is that Zotero is a citation manager first, while Open Paper is an AI reading assistant first, with citation management as a secondary capability. The repository has no GitHub releases, so there is no tagged stable version to pin against. Teams running self-hosted instances track the master branch directly. ## Conclusion Open Paper fits researchers who do most of their reading in PDF form and want AI answers tied to specific passages they can verify. The AGPL-3.0 license means any networked deployment must share modifications. Before self-hosting, confirm you have Postgres, S3-compatible storage, LLM API keys, and the patience to configure background workers; the README links to DEVELOPMENT.md as the actual starting point, not the main README itself. ## FAQ ### How do I use Open Paper AI? The hosted service at openpaper.ai requires only an account to start. For self-hosting, DEVELOPMENT.md in the repository is the starting point; it covers prerequisites including PostgreSQL, S3-compatible storage, LLM API keys, and Celery workers. ### What is Open Paper? Open Paper is an open-source research workbench for reading and annotating PDFs. It includes an AI assistant that cites every answer by linking to the exact passage in the source document, cross-paper querying, data extraction tables, Zotero import, and audio overviews. ### Is openpaper.ai free? The README does not document pricing for the hosted service at openpaper.ai. The open-source codebase is free to self-host under the AGPL-3.0 license, which requires sharing modifications when running a networked deployment. ## Sources - [Issues](https://github.com/khoj-ai/openpaper/issues) - [khoj-ai/openpaper on GitHub](https://github.com/khoj-ai/openpaper) - [License: AGPL-3.0](https://github.com/khoj-ai/openpaper/blob/master/LICENSE) - [Project website](https://openpaper.ai) - [README](https://github.com/khoj-ai/openpaper/blob/master/README.md) --- Hysen Labs editorial analysis, written from the project's own repository and release notes. Cite the canonical page: https://hysenlabs.com/projects/khoj-ai-openpaper