Agentic Publication Protocol (APP)
Motivation
In scientific research, the main product, the scientific paper, often contains only incomplete information about the work. Readers often need substantial effort to understand a paper and reproduce its results before using it in their own research. Important author know-how from the research process is often absent from the manuscript.
Recent progress in AI agents creates a new way to address this problem: instead of only publishing a paper, publish an agent. If every publication comes with an agent that can explain the paper, help readers reproduce the results, and even support follow-up work, each user can access a faithful and thorough representation of the scientific work. This can accelerate research and enable forms of scientific collaboration that were not previously possible.
What is APP
APP is an interactive format for authors to publish an AI agent together with research artifacts such as the paper, code, data, and related context. The bundle is a GitHub repository that users can open with any AI coding agent that reads AGENTS.md. The repo carries the paper alongside the code, data, and context needed for the agent to explain the work, reproduce figures, run experiments, and answer questions — more of what the research actually contains than a static manuscript can convey. The repo may also contain author-developed agent skills that are useful for understanding the work. A verified APP publication is a tagged public release with AGENTS.md plus an APP_PUBLICATION.json release manifest tying the release to validation and author approval. The structure of the published repository is illustrated below.
Readers clone the repo, open it in Claude Code, Codex, or any other agent that reads AGENTS.md, and the agent speaks for the paper.
Authors who plan to publish their paper in APP format can use the publish-paper skill. It orchestrates modular step skills that help authors reproduce/check existing results, organize publication materials, draft the AGENTS.md instructions, validate the publication, and carry out the final release or developer-sandbox outcome. The publish-paper workflow is shown below.
APP helps organize and reproduce/check results authors have already produced. It does not provide AI tools for improving the paper's scientific claims, adding new experiments, or carrying out new research.
This repository contains:
PROTOCOL.md — the specification of what an APP publication looks like.skills/ — official APP skills, including publish-paper, reproduce-results, prepare-staging, define-paper-agent, validate-publication, release-outcome, extract-chat-context, create-paper-page, and load-paper.template/ — starter files the skills adapt:
assets/readme/ — images used by this README..agents/, .claude-plugin/, and .codex-plugin/ — marketplace and plugin metadata for Codex and Claude Code.CONTRIBUTING.md — how to propose changes to the protocol, templates, and official skills.LICENSE — license terms for the protocol, skills, and templates.
Install and Update
First make sure the agent itself is installed. APP runs inside an AI coding agent — it is not a standalone program. Install Claude Code or Codex first, then follow the matching section below.
Claude Code
These are slash commands. Type them inside a running Claude Code session (at the Claude Code prompt), not in your shell. Press Enter after each line to install APP:
/plugin marketplace add LionSR/AgenticPublicationProtocol
/plugin install paper-protocol@paper-protocol
/reload-plugins
