Agentic Publication Protocol (APP)

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.

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:
  • template/AGENTS.md — starter for the publication's AGENTS.md.
  • template/README.md — starter for the publication's human-facing README.md.
  • template/CLAUDE.md — one-line Claude Code import (@AGENTS.md).
  • template/publications.md — template for the working repo's .publications.md release log.
• 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.

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.

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

If you haven't started a session yet, run claude in your terminal first, then enter the commands above. (/reload-plugins activates the plugin in the current session; the paper-protocol@paper-protocol form is plugin-name@marketplace-name, not a typo.) When it works, the plugin's skills such as /publish-paper become available — see Publish a paper for the next step.

To update an existing Claude Code install, type this inside Claude Code:

/plugin marketplace update paper-protocol
/reload-plugins

New skills, reference files, and templates appear after update; you do not need to reinstall the plugin.

These are shell commands. Run them in your terminal (your normal shell prompt), not inside Codex. First register the marketplace, then install the plugin:

codex plugin marketplace add LionSR/AgenticPublicationProtocol
codex plugin add paper-protocol@paper-protocol

codex plugin add may prompt you to authenticate on install. (paper-protocol@paper-protocol is plugin-name@marketplace-name, not a typo.) You can also install and toggle plugins interactively: open Codex, find Agentic Publication Protocol in the plugin browser, and press Space to enable it. Once enabled, its skills such as $publish-paper become available — see Publish a paper for the next step.

To update an existing Codex install, run this in your terminal:

codex plugin marketplace upgrade paper-protocol

New skills, reference files, and templates appear after update; you do not need to reinstall the plugin.

Use this path when you want to load the skills directly rather than installing the plugin. First clone this repo locally. Use HTTPS if you do not have GitHub SSH keys set up:

git clone https://github.com/LionSR/AgenticPublicationProtocol.git
cd AgenticPublicationProtocol

Or use SSH if your GitHub account is configured for SSH:

git clone git@github.com:LionSR/AgenticPublicationProtocol.git
cd AgenticPublicationProtocol

Then point your agent at the cloned skills/ directory.

To update a manual install, pull the latest changes in the cloned directory:

git pull

Then keep pointing your agent at the same skills/ directory. New skills, reference files, and templates appear after update.

The working repo is the folder that contains the materials you want to publish. It may also contain raw materials, notes, or other context that helps explain the paper.

You need an AI coding agent that can read APP skills and operate in your working repo. To complete publication, the required Git and GitHub operations must also be completed, such as creating commits, tags, releases, and release assets. You can either allow the agent to perform these operations or run them manually when the skill instructs you to do so. If you want the agent to perform them, you may need to set up GitHub authentication.

Open your working repo in an AI coding agent with this plugin installed, then invoke the publish-paper skill:

# Claude Code
/publish-paper

# Codex
$publish-paper

The publish-paper skill walks through the full workflow in order. It can span multiple sessions, and it may pause for author decisions between steps.

You can also call each step directly if you only need part of the workflow, want to resume from a known checkpoint, or are debugging a publication candidate. The step skills should be used in this order:

1. reproduce-results — reproduce/check existing paper results before staging, including figures, tables, experiments, and analytic derivations.
2. prepare-staging — organize the author-approved paper, code, data, environment notes, reproduction reports, and supplementary materials into publication-staging/.
3. define-paper-agent — draft and revise AGENTS.md, CLAUDE.md, and reader-facing documentation for the staged paper agent.
4. validate-publication — check APP structure, paths, privacy, consistency, reproduction status, and reader-agent usability.
5. release-outcome — perform the final review/freeze and either publish the validated release or record a developer-sandbox outcome.

publish-paper is still the recommended entry point because it keeps these steps coordinated and asks for author approval at the right checkpoints.

Clone the published paper repo and open it in an AI coding agent. Replace the placeholder URL below with the actual APP publication repository URL:

git clone https://github.com/author/their-paper.git
cd their-paper
# open in Claude Code, Codex, Cursor, or any agent that reads AGENTS.md

The agent reads AGENTS.md on startup and now speaks for the paper.

If the load-paper skill is installed, you can use it with a repo URL to clone or import a paper into the current project without leaving the working session.

The skills in this repository are grouped by how essential they are to the APP workflow.

Core publication workflow

Optional publication add-ons

Reader and import utility

This is a useful companion skill, but it is not required to create or validate an APP publication.

APP publications may point to reusable skills or extensions hosted outside the publication repo, including community-maintained skill collections or individual researchers' repositories. Use recommended_external_skills, app_extensions, or the External Skills and Extensions section of AGENTS.md to record the source URL, version or tag, and purpose. Skill sources should point directly to a directory containing SKILL.md, preferably at a stable tag or commit.

External skills are recommendations only. They are not part of an author-approved APP publication unless they are bundled into the tagged release and covered by the publication manifest.

Browse the growing list of papers published with APP in Discussion #36. Each entry links to the paper's release, its authors, and a summary of the work.

Published a paper with APP? Add it to the list in two steps:

1. Make sure your release is a fully validated, author-approved APP publication (it carries a valid APP_PUBLICATION.json).
2. Reply to Discussion #36 with a link to your release — any wording works. A bot verifies your release and adds it to the list automatically.

Contributions to the protocol, templates, documentation, validation behavior, and official APP workflow skills are welcome. Reusable field-specific or authoring skills can also live in independent repositories and be referenced from APP publications. See CONTRIBUTING.md for the boundary between official APP repo changes and external skill contributions.

PROTOCOL.md: CC-BY-4.0. Skills and templates: MIT.