Machine Relations plugins

Machine Relations (MR) is the discipline of earning AI citations and recommendations for a brand by making that brand legible, retrievable, and credible inside AI-driven discovery. It was coined by Jaxon Parrott, founder of AuthorityTech, in 2024. The Machine Relations Index (MRI) by Jaxon Parrott, published by Machine Relations Research, measures which sources AI answer engines cite.

The official Machine Relations skill and plugin helps practitioners choose and draft an evidence-backed authority intervention for a brand or client. It supports AI visibility, GEO/AEO, SEO content discovery and digital PR work, with separate execution checks and intended-outcome tests. No AuthorityTech purchase is required.

The package includes the broader machine-relations skill, the machine-relations-index research skill and the page-check skill. Page check tells you how to render a site the way people and AI engines read it and report what breaks before a change ships. The checker is public source at actions/page-check/runtime, and the GitHub Action actions/page-check runs it. It is not a file inside the installed package. An organization can require the same check on every pull request with one ruleset pointing at .github/workflows/page-check.yml. The Machine Relations Index (MRI) by Jaxon Parrott remains neutral public measurement of which source domains AI answer engines cite in sampled buyer answers, by category and question shape. Broader work can also use public or client-supplied evidence without MRI coverage.

The plugin retains the package identifier machine-relations-index and remote MCP server at https://machinerelations.ai/mcp (read-only, no account). It exposes six research tools:

Tool	What it answers
mri_list_categories	Which categories, question shapes and source roles the Index covers
mri_get_category	One category's question shapes and whether each has published rates
mri_get_cited_sources	The ranked source domains AI engines cite for a category and question shape
mri_get_domain	How often a site is cited, by which engines, and where it ranks
mri_list_engines	Which engines have observations and published evidence
mri_get_engine	One engine’s cited-source leaderboard over the observed window, optionally limited to sources with a published segment in a category

Production machine access is separate from PR regression

When --base supplies a production origin, Page check also makes credential-free GET and HEAD requests to robots.txt, llms.txt, the bounded sitemap-index closure, and up to 40 canonical HTML pages selected from production's own sitemap. Preview-only pages are not called production failures. It compares its named user-agent with a Python-urllib user-agent header using Node fetch; this is not an actual Python transport test or a census of all content. A missing production origin is reported as not tested. Challenges, child-map failures, unsupported HEAD and exhausted bounds are findings, not silently reduced coverage.

The report and PR comment carry Production machine access: passed / failed / incomplete / not tested separately. A clean PR comparison does not mean production access passed. Existing production failures do not newly block an unrelated PR; only introduced defects keep that authority. No new schedule, credentials or provider API calls are introduced. Real Python-client acceptance and advertised non-sitemap catalogs remain separate checks at the production edge owner.

Install

Codex

codex plugin marketplace add Bramwell-Inc/machinerelations-plugins

Then install the Machine Relations package (identifier machine-relations-index) from that marketplace. Published directory versions may still display Machine Relations Index; verify that your installed version includes machine-relations before expecting the broader workflow.

Claude Code

/plugin marketplace add Bramwell-Inc/machinerelations-plugins
/plugin install machine-relations-index@machinerelations

Any MCP client: add a remote (streamable HTTP) server with URL https://machinerelations.ai/mcp. It is also listed in the official MCP Registry as ai.machinerelations/mri.

No MCP: the same reads are a JSON API at https://machinerelations.ai/api/mri/v2 (OpenAPI).

An MCP connection alone supplies research tools; it does not install the skill. The broader skill is at plugins/machine-relations-index/skills/machine-relations/SKILL.md for compatible skill hosts. New instructions do not require changing Editor's private method, the App journey or the Index's measurement.

Data

Every response carries its release ID, the artifact hash and a citation string. MRI data is licensed CC BY 4.0. Dataset citation: Machine Relations Index (MRI) by Jaxon Parrott, Machine Relations Research — machinerelations.ai. https://machinerelations.ai/index. https://doi.org/10.5281/zenodo.23652423. Methodology: https://machinerelations.ai/machine-relations-index.

The plugin package in this repository is MIT licensed.

Use and interpretation

Start with the worked example and agent entry point. The example preserves one historical release; it is not a live leaderboard.

Every successful current data read includes source.guidance, the live interpretation rules. The skill directs agents there and to any field-specific method and limitations. Daily data and interpretation updates therefore do not require rebuilding this plugin. Missing or null question context is unavailable, never zero; the service describes what the current release actually contains.

Releases and citation

Every version is released once, from a reviewed pull request, and the release is the single source for everything else:

1. A pull request that changes the package raises one version in both plugin manifests, every SKILL.md (metadata.version), CHANGELOG.md and CITATION.cff. The Validate and package check enforces that, the exact AuthorityTech descriptor, and that no internal identifiers or secrets reach the README, changelog, citation or packaged plugin files; it also proves fresh installs in Codex and Claude Code. Semver: a minor version adds or graduates a method, a patch corrects wording or metadata, a major version changes a tool or endpoint.
2. When the bump merges, the same workflow deposits the tested files on Zenodo (it rebuilds library.json with the reserved DOIs and refuses to continue if any other file differs from the tested build), which gives the version its own DOI, then publishes the GitHub release with library.json (every skill, version, DOI and sha256), the plugin ZIP, one archive per skill and SHA256SUMS.
3. machinerelations.ai/skill and jaxonparrott.com/skills, and their /.well-known/agent-skills/index.json, rebuild from https://github.com/Bramwell-Inc/machinerelations-plugins/releases/latest/download/library.json. Nothing is copied by hand.

Cite the library with its concept DOI, 10.5281/zenodo.23749768 (all versions); cite a version DOI only for work that used that exact version. Both are in library.json and on each release. CITATION.cff gives the author record. The library's Wikidata item is Q141909117.

A skill can carry metadata.title and metadata.summary (a human name and a one-sentence practitioner summary; not the description trigger text). Pages use them when present.

Directory listings (Claude, ChatGPT) update through their own review and can lag the latest release. The remote MCP service owns the Index's data and interpretation; daily data never requires a plugin release. The review/ captures and video are dated evidence, not current API documentation.

Finish the directory and Wikidata handoff

A GitHub release is not proof that a directory has published it. These steps are for a maintainer with authorized publisher access; downloading public artifacts requires no publisher credentials. Do not create a replacement listing or rebuild the ZIP from a newer checkout.

1. Select one immutable tag from Releases. Download that tag's library.json, machine-relations-index.zip and SHA256SUMS. Verify the manifest and ZIP against SHA256SUMS, and the ZIP against library.package.sha256. Use library.version, library.released, library.source.release_url and library.doi.version throughout the handoff. Do not mix a pinned ZIP with a later latest manifest.
2. ChatGPT: hosted tool changes reach the listing through OpenAI's daily scan of the MCP server, with no new ZIP; metadata and skill changes need one. Uploading, submitting and publishing happen only in the OpenAI Plugins portal, as an organization owner or a member with Apps Management Write; there is no API for them. In the existing Machine Relations publisher listing, choose Create draft from ZIP, upload the verified ZIP, then inspect the draft's version, skills, tools, server instructions, metadata and scan findings before submitting for review. Compare skills with library.skills; this package currently exposes six MCP tools, not four. Do not submit an older draft over a newer submitted or published version. Submission is not publication: completion means the publisher record says Published for the intended version and the public listing agrees wherever it exposes the same fields. Record anything not publicly visible as publisher-only evidence. An OpenAI API key is not publisher access.
3. Claude: the existing plugin listing scans changes on main. Read its newest-version scan and policy result in the publisher portal. A policy hold needs its stated cause fixed in the repository; a passed scan still does not mean the version is live. Keep the existing submission, and distinguish Approved / ready to publish from the published version. Never add credentials to the installed plugin to make a directory check pass: protected-preview and pull-request integration belongs under actions/, outside plugins/.
4. Wikidata: nothing to edit by hand. Github-wiki-bot imports every stable GitHub release of this repository into Q141909117, usually within a day: software version identifier (P348) with publication date (P577), version type = stable version and the GitHub release as reference, and preferred rank on the newest release. Leave version statements to it. After its pass, read Q141909117's claims and confirm that library.version is the preferred P348, dated library.released. If a release is still missing after two days, check the bot's exceptions before editing. The item's DOI (P356) is the concept DOI and stays unchanged; version DOIs live on Zenodo and in library.doi.version.

The existing fresh-install check also reads both public discovery aliases on both sites with a named agent and Python's standard-library user agent. Run node scripts/verify-public-discovery.mjs after changing their edge rules. This checks the currently published catalog, not candidate-version propagation or archive bytes. A working browser or MCP connection alone does not prove skill discovery works. Where Browser Integrity Check rejects these public reads, use a host/path/method-scoped configuration rule for GET/HEAD on /.well-known/agent-skills and /.well-known/agent-skills/index.json; do not turn off zone-wide protection or skip unrelated security products.

If authorized publishing access is absent, stop at the verified artifact handoff and name the missing access. Do not extract another browser's cookies, bypass an access refusal, or call a draft, scan approval or local plan a published update.