Claude SEO — Run SEO From Your AI Assistant

Q: What stops the assistant from making a number up?

The tools return real rows with an as_of date and a named source, and the server's own instructions tell the model the rules it has to repeat: positions are period averages and not live ranks, and search volume comes only from the two keyword tools, fetched live from Google Keyword Planner and paid for with lookup credits. A model can still phrase something loosely, which is why every answer can be traced back to the tool that produced it.

The idea

What Claude SEO actually means here.

Most advice about using an AI assistant for SEO amounts to better prompting: paste a page in, ask for title ideas, paste the answer back. That works, and it is limited by the fact that the model cannot see your data. It does not know what you rank for, what moved last week, or which of your pages Google actually chose for a query — so it guesses, confidently.

An SEO MCP server changes the shape of that. The Model Context Protocol is an open standard for letting an assistant call tools, and MrKeyword implements it: 24 tools that read one workspace's real data. Connected, Claude stops guessing and starts querying. “Which of my pages is ranking for this?” becomes a lookup against your own Search Console rows, not a plausible sentence.

The honest limit of the idea is just as important. This is your assistant plus a data source. It is not a robot that rewrites your site overnight, and MrKeyword does not run one.

Setup

Two minutes, three steps.

You need a MrKeyword account — Free is enough — and a client that speaks MCP.

Create a token
In MrKeyword, open Settings → Connections and find API & MCP Tokens. Give it a label and press Create. Leave “Allow this token to make changes” unticked and the token can only read. It is shown once — only its hash is stored.
Point Claude at it
For Claude Code, one command. For the Claude desktop app it is claude_desktop_config.json under Settings → Developer → Edit Config; for Cursor it is .cursor/mcp.json. The developer reference has all three.
```
claude mcp add --transport http mrkeyword \
  https://app.mrkeyword.io/mcp \
  --header "Authorization: Bearer mk_your_token"
```
Ask it something
Try “what is hurting my site today?”. Your client will list the tools it can see, and that list is always the authority on what your workspace can do.

Ask your AI · Example

Ten things to ask once it is connected.

Each of these is answered by the tools named under it, out of your own Search Console data and MrKeyword's own crawler. None of them needs a token that can make changes.

Which of my pages is Google actually ranking for “emergency plumber near me”?
Tools usedget_page_keywords
What is hurting this site today, and what should I do first?
Tools usedget_daily_issues
How has search performance moved since last month?
Tools usedget_site_summary
Which keywords dropped this week, and by how much?
Tools usedget_rank_summaryget_rankings
Show me the queries I am on page two for — the ones worth one afternoon.
Tools usedget_search_queries
What did the last crawl find, grouped by how many pages it affects?
Tools usedget_site_auditget_audit_issues
Are any of my programmatic templates thin or near-duplicate?
Tools usedget_page_templates
How many people search for “emergency plumber” in the UK each month? (uses one keyword lookup)
Tools usedresearch_keyword
Summarise last month for the client, with the sources.
Tools usedget_site_summaryget_report

Example questions, not recorded answers. Every tool named here is one the MrKeyword MCP server answers today — see the full reference.

Prompts

Four jobs already written.

The server exposes MCP prompts as well as tools, so a client that lists them gives you these as commands. Each drives several tools and comes back with the answer written up rather than a table dumped.

weekly_seo_review

Weekly rank review for one site: what moved, what needs attention, and five actions for this week.

fix_my_top_issues

Walk the latest site crawl: what is broken, grouped by category, with concrete fix steps per issue.

find_keyword_opportunities

Turn near-miss Search Console queries into a prioritised list, sized with live Google Keyword Planner volume when you agree to spend lookups.

client_report

Write a client-ready SEO report for one site from rankings, search queries and the latest site audit.

Clients

Which assistants this works with.

Claude Code and the Claude desktop app are what we develop against; both take an HTTP MCP server with a bearer header. Cursor uses the same JSON shape in .cursor/mcp.json. Anything else that implements the Model Context Protocol over HTTP will work, because there is nothing bespoke in the implementation: it is JSON-RPC 2.0 over a single POST, with no session to keep alive.

We do not list ChatGPT as supported. MCP support in consumer chat clients moves quickly, and a tool page that claims compatibility it cannot test is how people waste an afternoon. If your client can add an HTTP MCP server with a custom header, MrKeyword will answer it; check your client's own documentation rather than ours.

If you would rather not use MCP at all, the same mk_ token authenticates the REST endpoints under /api, so a script can do the same work over plain HTTP.

Scope

What this does not do

An assistant with access to your data should come with a short list of what it cannot do.

It is not an autonomous agent. Nothing here edits your website, publishes content or changes your Google or Bing accounts.
Read-only by default. Five tools can change something inside MrKeyword, and only for a token created with “Allow this token to make changes” ticked.
One workspace per token. Anything belonging to another workspace answers as though it does not exist.
It cannot see competitors. The tools read properties you have connected, so there is no competitor data to fetch.
The model can still be wrong. Every figure it quotes came from a tool with a named source and an as-of date — which is exactly what lets you check it.

Questions about using Claude for SEO

Does this work with ChatGPT?

It works with any client that speaks the Model Context Protocol over HTTP. We test Claude Code, the Claude desktop app and Cursor, and the same URL and bearer token work from a script or plain curl. We are not going to list ChatGPT as supported, because what its clients support changes and we would rather you checked your own client than took our word for it.

Do I need to be a developer?

No. You create a token in Settings → Connections, paste one line into your assistant, and ask questions in English. The developer reference exists for people who want to call the endpoint directly, not because you have to.

What can it actually read?

One workspace: its websites, Search Console performance, tracked keyword positions and history, the queries you appear for, crawl results and audit issues, saved reports, content briefs and your plan usage. 24 tools in total. It cannot read anything outside that workspace.

Can Claude change my site?

No. It cannot touch your website, your Google account or your billing under any setting. Within MrKeyword, a token is read-only unless you tick “Allow this token to make changes” when you create it, which unlocks five tools — tracking keywords, setting a target page, annotating, starting a crawl and running a saved report. Every one of those is written to the workspace audit log.

Is this an autonomous SEO agent?

No, and it is worth being precise. There is no MrKeyword process that goes off and edits your site. What exists is a data source your assistant can query, plus tools that can queue work you asked for. The judgement and the publishing stay with you.

Does it cost extra?

No. Connecting an assistant has no separate charge on any plan, including Free. Two tools use the workspace's allowances — tracking keywords and starting a crawl — exactly as they would in the app. Fair-use limits apply at 120 calls a minute per token.

What stops the assistant from making a number up?

The tools return real rows with an as_of date and a named source, and the server's own instructions tell the model the rules it has to repeat: positions are period averages and not live ranks, and search volume comes only from the two keyword tools, fetched live from Google Keyword Planner and paid for with lookup credits. A model can still phrase something loosely, which is why every answer can be traced back to the tool that produced it.

Go deeper.

MCP reference

All 24 tools, scopes, rate limits and a snippet for each client.

AI SEO

What AI does inside the product: explaining issues and drafting the fix.

GEO and AEO

Getting cited by AI answers — what is known, and what nobody can measure yet.

Pricing

MCP access is included on every plan, including Free.

Connect your assistant.

Create a token, paste one line, and ask it anything. Free forever, no credit card.

Get started free No credit card required.

Claude SEO: ask your assistant, not a dashboard.