The RankForge agent, inside your editor

An audit report tells you what's wrong; your coding assistant can actually change the code. With RankForge connected over MCP, one prompt (“audit my dev server and fix what you find”) runs the full technical-SEO audit against localhost, walks the findings, patches your files, and re-audits to verify. The measurements are deterministic (no AI guessing); the reasoning is your assistant's.

Install the plugin. It brings the MCP tools and two skills:

/plugin marketplace add Barbudda/rankforge-cli
/plugin install rankforge@rankforge

/rankforge:seo audits the site in your repository and fixes what it can; /rankforge:check compares your local or preview version with production before you ship. Or add the bare MCP server:

claude mcp add rankforge -- npx -y rankforge-cli mcp

On native Windows, use cmd /c npx -y rankforge-cli mcp after the --. Then ask: “Use rankforge to audit http://localhost:3000 and fix the issues you can.”

Add a stdio server to your client's configuration (e.g. .cursor/mcp.json; VS Code's .vscode/mcp.json uses a servers key instead):

{
  "mcpServers": {
    "rankforge": { "command": "npx", "args": ["-y", "rankforge-cli", "mcp"] }
  }
}

It needs Node.js 18 or later. The README has the exact file for each editor.

• audit_site: a full deterministic audit with category scores, every issue with evidence, orphan pages, click depth, duplicate-content clusters, semantic internal-link suggestions, measured image weights and broken links. Works on localhost.
• compare_sites: production against your dev server or a preview deployment, page by page. It reports what would drop pages from search (new noindex, pages that start failing, robots.txt blocking Google) and keeps differences that come from the preview environment apart.
• audit_page: an instant single-page check while you iterate.
• geo_audit and generate_llms_txt: whether AI search engines can reach, read and cite the site, and an llms.txt built from your pages.
• get_fix_template: a framework-idiomatic patch (unified diff) for mechanical issues: robots, sitemap, viewport, lang, canonical, structured data, OpenGraph image.
• seo_docs: grounded answers from the RankForge knowledge base, so the assistant explains issues accurately.
• list_rules: the 54 checks, under the same rule ids the audits return.

The MCP server is the open-source CLI, started by your editor. It runs on your machine, so it audits localhost and you see issues as you build, before deploying. It needs no account and no key, and it sends nothing anywhere: the only outside call is the optional Core Web Vitals lookup, which asks Google about the URL when you request it.

The tools only read rendered pages, the same thing any crawler sees. RankForge never writes to your repository: your assistant edits the files, and you review them as you would any change. The crawler identifies itself as RankForgeBot and honours robots.txt unless you pass ignoreRobots for your own site. The hosted endpoint at /api/mcp is closed by default and, when opened, refuses private and internal addresses, checking every resolved IP, redirects included.