Claude Code plugin

The plugin puts the Sitearc MCP tools inside a Claude Code session. It adds eleven skills that guide a full workflow. A skill is the difference between a tool call and a piece of work that ends somewhere useful.

Install

/plugin marketplace add ethanplusai/sitearc-app
/plugin install sitearc

The marketplace is available to account holders while the repository is private.

The first tool call opens an OAuth prompt. Sign in at app.sitearc.ai and approve the connection. You need a Sitearc account.

Check it worked by asking for the brief:

Read the project brief for my site.

The eleven skills

Every skill takes a projectId. Ask Claude to list your projects if you do not know yours.

seo-project-setup

Run this first, once per site. It interviews you about scope, goals, positioning, competitors and key pages. The answers become that project’s shared context.

Every other skill reads that context. It survives new sessions, new machines and new agents, so you answer once.

seo-audit

Use this when you want a shareable report a non-expert can act on. It produces one page in plain language, built around a single action to take this week.

Reach for it when a client or a colleague asks what is wrong with the site.

keyword-research

Use this to turn seed topics into a prioritised opportunity set. It pulls metrics, reads the result pages, and saves the terms worth keeping.

keyword-clustering

Use this after research, when you have terms and need pages. It groups keywords by intent and maps each cluster to an existing or proposed page.

This is keyword mapping, not just semantic grouping. The output is a page plan.

competitor-analysis

Use this for one named competitor. It works through their organic footprint, ranking keywords, content themes, backlinks and gaps.

The question it answers is what to learn from, counter-position against, or outrank.

competitive-landscape

Use this for the market rather than one rival. It maps the leaders, the content themes that are working, keyword coverage and the openings.

Run it before competitor-analysis when you do not yet know who the leaders are.

link-prospecting

Use this when you have something worth linking to. It finds realistic pages, sites and authors, then works out contact paths and drafts outreach.

Prospect discovery uses Sitearc data. Contact discovery uses your agent’s own web and browser tools.

local-seo

Use this when rankings depend on a place. It audits the Business Profile, compares it to local competitors, and maps Maps visibility around a location.

For national organic work use keyword-research or competitor-analysis instead.

fanout-research

Use this when the question is about assistants rather than search results. It researches the sub-questions an assistant fans out to. Then it turns those into retitles and expansions of pages you already have.

It is a research skill, not a publishing one. It will not generate a page per question, and that restraint is the point of it.

seo-coach

Use this when you are new and want a guide rather than a workflow. It explains what each workflow does and recommends a next step.

seo-loop

Use this to run the measured cycle. Brief, evidence, a proposal with a prediction, then a person’s approval. Then the pull request, the preview validation, the window, the verdict and the learning.

It is also the right skill for asking what has actually been proved about a site. Its honesty rules are below, and they are the reason the answer is worth trusting.

Skills being built

Five more are designed and not shipped.

• Review a draft and fix or flag issues before it publishes.
• Apply an approved recommendation through WordPress or a pull request.
• Turn open feedback comments into recommendations.
• Propose a patch to a site’s brand profile.
• Read everything the platform knows about a site into context.

The honesty rules, in words

The seo-loop skill carries a block of rules that constrain what the agent is allowed to say. They contain no digits at all, on purpose. An illustrative number inside an anti-fabrication rule anchors a fabricated number in the output.

Start with the brief. The first call every session is get_project_brief. The reply opens by stating what is connected and what is not. It names the one concrete thing that would change that.

Trust the tool’s own provenance. A result the tool did not mark as failed is real data for that project. The agent reports it at the tier and source it names. It does not append doubt because a title looks generic.

Doubt what the tool doubts. Skepticism belongs to a result marked failed and to a source reported as not connected. Those are the cases where doubt is the behaviour asked for.

A tier is only as good as its citation. A learning that cites no experiment prints nothing where the citation would be. That silence is the signal. A measured tier with no citation is on file and unproved, not shown.

No number the tools did not return. The agent types no digit next to a metric unit unless a tool returned that digit. There is no exception for illustration, and none for a hypothetical.

No arithmetic on measurements. A lift, a delta, a percentage change or a per-day rate is a number no tool returned. Concluding an experiment is a tool’s job, not arithmetic in a reply.

A refusal is the answer. When a tool returns inconclusive or refuses, the reply ends at what it returned and why. No band, no gap, no directional read, no promising trend after the word.

Never paraphrase a standard upward. A verdict is quoted or summarised downward. “The change worked” is not a paraphrase of a verdict that did not say so.

No guessed rankings. Naming the pages a change touches is fine. Ordering them by expected effect is a ballpark by structure. That holds in a table and in prose, and it is not allowed.

Hedges do not launder a number. Writing roughly, ballpark, directional or a tilde before something that reads like a measurement does not make it safe.

Approval belongs to a person. No tool approves a proposal. If ship_experiment refuses, the agent relays the refusal in the tool’s own terms and stops. Suggesting a way around it is the worst failure available.

Use the tools, and only the tools. No direct database queries, no direct calls to a data provider or a Google API, and no suggesting either. That holds even when a tool fails.

Every report carries its sources. A source-and-tier column on every data table, per row. A provenance footer naming the tiers used and the window. It states what was collected against what was expected. It names any overlapping confounder, and the one thing that would raise the weakest tier.

Try it

• “Research keywords for my website and shortlist the best opportunities.”
• “Audit my website and tell me what to fix first.”
• “Which pages are close to ranking in Google Search Console?”
• “What has this project actually proved, and at what tier?”