Repo Shield

Claim your code. Stop AI pipelines from silently training on it.

Public repos get bulk-ingested into AI training corpora every day — no consent, no credit, often no trace. Repo Shield gives you a simple, honest, enforceable position: a signed LICENSE, an explicit AI-training NOTICE, a machine-readable policy, and a weekly watch that keeps them from fading.

Free, open source, and only as complicated as you want it to be. Web: repo-shield · CLI: @reposell/repo-shield · Agent skill: skills/repo-shield/SKILL.md

Honest caveat: this is deterrence, not enforcement. Generated files stake your position and make accidental ingestion visible. Nobody can "block" AI scrapers, and anyone who promises that is selling a lie.

Table of contents

• Features
• Quickstart
• CLI reference
• For AI coding agents
• What's in the pack
• Pick a license
• No GitHub account?
• What Repo Shield isn't
• Open source
• Contributing & support
• Development

Features

• Signed license, canonical text. The complete official SPDX text for all five licenses is embedded in your repo — no truncated copy, no paste-in-the- notes approximation. MIT and ISC get your holder and year filled in.
• Explicit AI-training NOTICE. A short legal claim that names what is and isn't allowed, with an SPDX identifier the whole pack stays matched to.
• Machine-readable policy. AI_TRAINING_POLICY.md lets scrapers and tooling check consent without parsing prose.
• A self-healing watch. A weekly GitHub Actions check runs inside your repo with your own token — Repo Shield never sees your code — and opens an issue the moment a protected file is removed, edited, or a license id drifts.
• One review, one PR, per repo. Nothing touches your default branch until you merge. Re-running just reuses the open PR.
• Agent-ready. The exact same flow is packaged as an Agent Skill, so an AI coding agent can protect owner/a owner/b or --all with the one binary a human runs. See For AI coding agents.

Quickstart

Requirements: Node 18+, and the Repo Shield GitHub App installed on the repos you want to protect. Three commands:

npm i -g @reposell/repo-shield   # installs `rs` (and `repo-shield`)
rs login                    # device-flow sign-in — code in terminal, no password
rs protect --all            # one protection PR on every repo the app can write to

Or protect a single repo: rs protect owner/repo.

Install the GitHub App on the repos you want to protect — yours, your org's, or all of them. Any GitHub user can install it; the app is the write mechanism rs uses to open pull requests on your behalf: https://github.com/apps/reposell/installations/new

What happens next: Repo Shield writes the pack as one commit on a repo-shield/protect branch and opens one pull request per repo. Nothing on your default branch changes until you review and merge. Running it again on the same repo just reuses the open PR.

CLI reference

| Command | What it does | |---------|--------------| | rs login | Device-flow sign-in. GitHub prints a code, you approve it in the browser; a token is stored in ~/.config/repo-shield/config.json (0600) on your machine. | | rs whoami | Your account + whether the app is installed. | | rs repos | Every repository the app can write to. | | rs protect owner/repo | Opens one protection PR on that repo. | | rs protect owner/a owner/b | One PR per repo, up to 50 per run. | | rs protect --all | One PR per repo across every installed repo. | | rs logout | Forgets your session. |

rs protect options:

--license <id>   mit | isc | unlicense | apache-2.0 | gpl-3.0   (default: mit)
--holder <name>  Copyright holder (default: your GitHub name)
--year <n>       Year of first publication (default: current year)

Protected files are written as one commit on a repo-shield/protect branch. If a PR for the repo is already open, rs protect reuses it.

For AI coding agents

Repo Shield is a first-class tool for coding agents. The same flow a maintainer runs by hand is packaged as an Agent Skill — an agent can be pointed at skills/repo-shield/SKILL.md, sign in once, protect one repo or an entire organization, and hand the PR links back. No separate API, no server, no admin.

npm i -g @reposell/repo-shield
rs login
rs protect owner/a owner/b --license isc

Why it's safe to hand to an agent:

• Same CLI, no second system. There is no web API or different tool for the agent to learn — it drives the exact rs binary a human runs.
• Org-wide in one command. protect owner/a owner/b, or --all for everything the app is installed on.
• Reviewable by construction. Every repo gets its own pull request on a repo-shield/protect branch. Nothing is ever pushed to a default branch without a human merging it.
• It says when it can't. A 403 or missing-install surfaces as a clear re-run step; the skill does not skip repos silently.

Install the skill

The skill is a single file at skills/repo-shield/SKILL.md — instructions plus the exact agent protocol (what to run, how to read the output, what to report). Any agent that accepts skill files can load it:

• Claude Code — copy the folder into ~/.claude/skills/repo-shield/
• Cursor — copy into .cursor/skills/repo-shield/
• Codex / Copilot / any agent — open SKILL.md and follow it

Or simply tell your agent to read the file before working on the repo:

Read skills/repo-shield/SKILL.md and follow its protocol to protect this repo from AI training.

What's in the pack

| File | Purpose | |------|---------| | LICENSE | Your chosen license, with an SPDX identifier | | NOTICE | Copyright + explicit AI-training consent clause | | AI_TRAINING_POLICY.md | Machine-readable consent key |

Plus two kept-in-sync extras for public repos:

| File | Purpose | |------|---------| | REPO_SHIELD.txt | Signature marker that flags copies in scrape scans | | .github/workflows/repo-shield.yml | Weekly monitor (runs in your repo, not ours) | | scripts/shield-check.mjs | The check itself — verifies files, license, markers |

The monitor runs weekly on GitHub Actions with your repo's existing GITHUB_TOKEN. Repo Shield never sees your code.

Pick a license

The defaults are the safest, most common choices for public code:

| License | When to use it | |---------|----------------| | MIT | The default. Permissive, tiny, understood everywhere. | | ISC | Minimal permissive — the whole grant fits in a paragraph. | | Apache-2.0 | Permissive plus an explicit patent grant, for big projects. | | GPL-3.0 | Copyleft: derivatives must stay free. Filed as GPL-3.0-only. | | Unlicense | Your code, public domain. |

Every license is written into your repo as the complete canonical text from the official SPDX registry — including Apache-2.0 and GPL-3.0, no paste step needed. MIT and ISC get your holder and year filled in. An approximate license is worse than none.

Every generated LICENSE carries an SPDX-License-Identifier header, and the monitor verifies the identifier is from the known set and matches the NOTICE. No AI-training clause, or a wrong license id, fails the weekly check and opens an issue telling you exactly which file to fix.

No GitHub account?

Build the pack straight from the manual mode on the site (#/generate) — pick the files you want and download them, or use apply-all.sh offline.

What Repo Shield isn't

• Not enforcement. No tool can block AI scrapers. These files make the position explicit and the removal visible.
• Not a watermark. No banner, sticker, or altered source — the repo you publish is the repo you own.
• Not a scanner. Repo Shield protects your repos; it does not scan or police what others do with their code.

Open source

Repo Shield is free and open source, with the condition that it stays yours: the Repo Shield Source License (No AI Training) lets anyone use, read, and improve this code, and explicitly forbids using the software itself to train a machine-learning model. Full terms in LICENSE.

If the tool saves your repos from a scrape, a small donation covers hosting and maintenance:

• Donate
• Star on GitHub
• Read the design system

Contributing & support

Issues and pull requests are welcome — this product runs on the gates that make it trustworthy, so fixes should keep every gate green:

• Open an issue for a bug or a suggestion
• Read DESIGN.md before changing UI or tokens
• Run the quality gates before opening a PR (see below)
• For hosting and maintenance donations, use the support page

Development

npm install
npm run dev        # local dev server
npm run build      # production build
npm run typecheck
npm test           # pack + license + CLI protect-flow unit tests

Static Vite + TypeScript app, deployed to GitHub Pages. No server components. The published CLI is compiled JS and runs on Node 18+; running the TypeScript sources directly (tests, npm run cli) needs Node 22.6+.

Quality gates, all run in CI on every push/PR (.github/workflows/ci.yml):

• python3 scripts/check_no_emoji.py — zero emoji in UI, taste docs, or the agent instruction surface
• python3 scripts/validate_tokens.py — token files parse and aliases resolve
• python3 scripts/validate_contrast.py — token pairs pass WCAG, light + dark
• python3 scripts/validate_component_spec.py — component specs are complete
• node .impeccable/repo-shield-verify.mjs all — real-render gate in Google Chrome: computed-style contrast + overflow across all six routes at 1440px and 280/320/414px. Expect PASS: contrast + overflow, 6 route(s) clean. Needs npm run build, then npm run preview running on :5178, and Chrome installed.
• npm test / npm run typecheck / npm run build

Releasing the CLI

The npm package (@reposell/repo-shield) auto-publishes from GitHub Actions when a v* tag is pushed, using trusted publishing (OIDC) — no token needed locally. Connect the package to this repo once in npm settings: package > Settings > Set up trusted publishing, GitHub, owner EnzoVezzaro, repo repo-shield, workflow filename release.yml, allowed action npm publish (the workflow already carries id-token: write).

# bump cli/package.json first (tag must match the version exactly)
git tag v0.1.6 && git push origin v0.1.6