You can use Claude Code for SEO by giving it a repeatable skill, a website it can inspect, and a precise output contract. The popular AgriciDaniel/claude-seo project is a community plugin, not an Anthropic SEO product. Its commands cover page audits, technical checks, content, schema and AI-search readiness. The useful question is which findings you can prove and fix.
This guide was checked on November 12, 2026. We reviewed installation documentation and built a small synthetic triage example. We did not run a live competitive benchmark between the community plugin and Orkas, and no private site’s traffic data appears here.
Skill, MCP and SEO data solve different problems
| Component | What it contributes | What it does not establish |
|---|---|---|
| SEO skill | Instructions, checklists, scripts and report structure | Access to your accounts or paid databases |
| MCP connector | Tools an agent can call with the service permissions you grant | A complete SEO strategy, or authority to publish changes |
| GSC and analytics | Your measured visibility, clicks and visits within their reporting limits | Total market search volume or competitors’ complete backlink profiles |
| Public HTML and browser | Status, metadata, links and a rendered page | Proof that Google indexed the current version or that users converted |
If you already work in Claude Code and want a reusable checklist, a community skill is a natural starting point. If you want a finished audit from a desktop conversation, Orkas’s built-in SeoGeoAgent is another route. If you only have a handful of files, a well-scoped brief can be enough before installing anything.
Install the community plugin in Claude Code
The project’s current installation guide recommends these commands inside Claude Code, not in your operating system’s shell:
/plugin marketplace add AgriciDaniel/claude-seo
/plugin install claude-seo@agricidaniel-claude-seo
/seo setup
/seo doctorPlugin installation and runtime provisioning are separate. The project lists Python 3.11–3.14, Git and a configured Claude Code CLI; its managed setup prepares an isolated environment and Chromium. Browser-dependent checks need that browser to work. A plugin appearing in the list is not proof the crawler is ready. Use /seo doctor and read degraded or failed checks before proceeding.
Review the repository, scripts, permissions and intended output before installing. Optional data extensions have their own keys and costs. “Free skill” means its code is free to use under its license; your Claude usage, external APIs and infrastructure can still cost money.
Start with one page, then widen the crawl
/seo page https://example.com/Replace the example URL with a page you own. Begin with the homepage and one revenue-relevant page rather than asking for an unlimited crawl. In your request, define the intended audience and action, so the audit does not recommend a generic blog expansion for a shop that needs better product pages.
Audit these two URLs for search and AI-answer readability. Return at most five prioritized issues. For each, give the URL, observed evidence, why it matters, the smallest proposed repair and a way to verify it. Separate confirmed defects from hypotheses and unavailable data. Do not edit, submit or publish anything.
That is an example brief, not a claim that it was run on example.com. A useful report should include an inspected URL list and crawl limits. If rendering fails, it should distinguish unavailable rendered evidence from the HTML it could retrieve.
A synthetic example: turn observations into a repair queue
Illustrative shop page: imagine a product page with two main headings, a canonical pointing to a staging domain and a title that is also used by the homepage. These facts require different decisions; they should not all receive the same “critical” label.
| Observation | What to establish | Proposed next step |
|---|---|---|
| Canonical points to staging | The intended public URL and any deliberate duplication strategy | Correct the canonical if it is wrong; inspect the served response and later Google’s selected canonical |
| Two H1 headings | Whether the page’s main purpose is unclear, not just an arbitrary count | Make the product heading and section hierarchy clear; check visual and raw HTML output |
| Title duplicates the homepage | Which product intent this page is meant to satisfy | Write a specific, accurate title and measure query-level changes after publication |
No traffic increase can be inferred from this synthetic example. An audit’s score is a tool-specific prioritization aid, not Google’s ranking score. The repair queue is useful when someone can validate its evidence and accept or reject the proposed change.
Connect Search Console only when the question needs it
Public inspection can find broken canonicals without account access. To decide which existing page has valuable impressions but few clicks, add GSC through an authorized connector or an export. Keep property, date range, search type and filters with every number. Query rows can omit anonymized searches, so their sum is not necessarily the page total.
GSC is not a keyword-volume database. A skill does not create missing market data: if search volume, links or AI referrals were not retrieved, the report must say so. Our Google Search Console MCP guide explains API setup and limits. For GEO, inspect actual cited answers and referrals separately; a clean llms.txt file does not guarantee citation.
Close the loop after a repair
- Save the original response or file and the exact proposed diff.
- Check the repair in a local preview: visible answer, metadata, links and any applicable language versions.
- Publish through your normal approval process, then fetch the public URL again. A local change alone is not a served repair.
- Compare matched periods and queries later; annotate other launches, seasonality and indexing changes.
Do not automatically create large numbers of pages from the audit. An independent intent needs a useful answer and a believable next step. Similar wording alone does not prove two pages should merge, and a single search-results overlap threshold does not guarantee unchanged rankings or AI citations.
Using Orkas instead of assembling a skill stack
In Orkas, ask the built-in SeoGeoAgent for a scoped audit and connect GSC only if it needs your property data. You can also run your installed Claude Code with a community skill in a project folder. These are alternative setups, not a measured claim that one finds more problems. Ask for the same evidence, missing-data labels and repair checks either way.
See the agentic SEO workflow for the broader process, or SEO/GEO work in Orkas for the task entry.
FAQ
Is claude-seo an official Anthropic product?
No. AgriciDaniel/claude-seo is a community project. Claude Code’s skill and plugin mechanisms are official, but that does not make every plugin an Anthropic service.
Does installing a skill connect my GSC account?
No. Installation, account authorization and data retrieval are separate steps. Supply an authorized connector or export when the audit needs private property data.
Will the skill improve rankings automatically?
No. Findings still require verification, prioritization, implementation and measurement. Neither an audit score nor a generated report guarantees search traffic or AI citations.
What if the browser part fails?
Read the readiness report and distinguish raw HTML checks from rendering checks. Fix the managed runtime before claiming screenshots, visual results or JavaScript-rendered content were inspected.
Sources and scope
Community repository, its installation guide, Anthropic skill documentation and Google’s guidance on generative AI content, checked November 12, 2026. Installation commands were checked against the source documentation, not executed here; no side-by-side agent performance test was run.