SEO Toolkit

Eight SEO and AI-search workflows for Claude and ChatGPT, built on measured data rather than invented numbers. MIT licensed.

Install

In claude.ai or the Claude desktop app, go to Customize > Plugins > Add > Add marketplace and enter:

Bishwas-py/ai-seo-toolkit

Then select AI SEO Toolkit and Add.

From Claude Code instead:

/plugin marketplace add Bishwas-py/ai-seo-toolkit
/plugin install ai-seo-toolkit@webmatrices-seo

That is the whole install. It covers chat on web, desktop and mobile, Cowork, and Claude Code from one account. To pick up a new version later, run /plugin update ai-seo-toolkit@webmatrices-seo, or turn on auto-update for the marketplace.

Other clients

The plugin carries two things, and they reach different places. The skill is the routing and the house rules, and it loads in Claude only. The MCP server carries each workflow's full procedure in the tool's own response, so it stands on its own anywhere it can run.

Claude Desktop, by hand. The bundled server runs in Cowork and Claude Code, not in a plain desktop chat. To get the workflows there, add it to ~/Library/Application Support/Claude/claude_desktop_config.json:

{
  "mcpServers": {
    "ai-seo": { "command": "npx", "args": ["-y", "ai-seo-toolkit"] }
  }
}

Codex. Either codex mcp add ai-seo --command npx --args -y ai-seo-toolkit, or write it into ~/.codex/config.toml yourself:

[mcp_servers.ai-seo]
command = "npx"
args = ["-y", "ai-seo-toolkit"]

The skill does not load in Codex, so no workflow is chosen for you. Call the one you want by name and follow what it returns.

ChatGPT. Not supported, and not through any flag. ChatGPT reaches MCP servers over HTTPS and never launches one on your machine, so a local stdio server is out of reach. Hosting this behind an HTTPS endpoint or OpenAI's Secure MCP Tunnel would work, and would also put the browsing and the Search Console session on a server rather than your own signed-in Chrome, which is most of why the numbers here are measured. The Chrome extension is the honest answer for ChatGPT.

Related MCP server: GSC-MCP

Pair with MCP Browser

Every workflow that touches a URL must fetch the page and prove it read it. A plain HTTP fetch cannot distinguish a genuinely thin page from one that only looks empty to a fetcher, and that distinction is the entire diagnosis when a site renders through JavaScript.

MCP Browser drives real Chrome with your signed-in session, so the toolkit can render JavaScript, fetch both ways and compare, and read Reddit, BlackHatWorld and forums behind a login.

brew install --cask bishwas-py/tap/mcpbrowser

macOS on Apple Silicon, free for 50 requests a day. Everything here works without it, and will say in one line when it could not verify rendering.

What it refuses to recommend

Google publishes an AI optimisation guide that contradicts most of what is sold as GEO and AEO. These workflows encode the guide. They will not tell you to add an llms.txt file, chunk your content, write in a special style for AI, treat structured data as a citation requirement, or buy brand mentions, and they will push back if asked.

Two findings they do hold to: a page must be indexed and eligible to show with a snippet before it can appear in a generative answer, and across 1.4 million prompts the title, snippet and URL decided whether a page was opened at all, while the average cited page was around 500 days old.

Develop

npm install
npm test                             # drives the real server over stdio
npm start                            # run the server directly
claude --plugin-dir .                # load the plugin from this working copy
claude plugin validate .             # check the manifest and components

server/workflows.js is the single source of truth. Each workflow declares its arguments once, and the server derives both the MCP prompt and the MCP tool from that declaration, so the two can never drift apart.

Release

Pushing a tag is the whole release.

npm version patch && git push --follow-tags

That bumps package.json, mirrors the number into .claude-plugin/plugin.json through the version lifecycle hook, commits, tags, and pushes. CI then runs the tests, refuses to continue if the tag disagrees with package.json, publishes the server to npm and cuts a GitHub release with generated notes.

Installed plugins pick the new version up from this repository on their own. The npm package is what .mcp.json starts, so it is plumbing rather than something anyone installs by hand.

There is no NPM_TOKEN. The workflow authenticates with npm trusted publishing, which exchanges a short-lived GitHub OIDC token for publish rights and attaches provenance automatically, so there is no long-lived credential to leak or rotate.

One-time setup

Trusted publishing is configured on an existing package, so the first publish is manual:

npm login
npm publish            # claims the name

Then on npmjs.com, under the package's Settings, add a trusted publisher: repository Bishwas-py/ai-seo-toolkit, workflow release.yml. Every release after that is the one-liner above.

The repository must also be public before the release asset is downloadable by anyone else. While it is private, the download link on the landing page returns 404 for everyone but you.

The test suite asserts the MCP surface over a real stdio connection and checks the content guarantees, including that no workflow recommends a tactic Google has said does not work.