Free SEO MCP that closes the loop: find, fix, ship as a PR or to WordPress, prove. 50 tools, no key.
Most SEO tools give you a report. OnPage.dev gives your assistant something to do: write the fix, check it before you publish, and confirm it worked. It also checks what AI search sees, not only Google: which AI crawlers may read your site, whether your content is ready to be quoted, and whether you have an llms.txt.
You: Fix the SEO on our pricing page and keep going until it scores 90 or more.
Assistant: scan_page 46/100 no meta description, no H1, title 14 characters
get_fix_pack title, description and canonical, written from the page
generate_schema Product JSON-LD, 2 values to fill in
(edits pricing.html)
scan_html 92/100 before deploying, all errors gone
rescan_and_compare 94/100 live, +48 since the first scan
Done. 46 to 94. Two prices in the JSON-LD still need your real values.
Already connected Google Search Console, GA4, Ahrefs or Semrush to your assistant? OnPage.dev tells the assistant to pull their numbers in, then prioritize_fixes ranks every fix by the traffic it can win. Nothing connected? It ranks by severity, as before. You do not set anything up.
You: Audit example.com and tell me what to fix first.
Assistant: start_site_audit 10 pages, average 81
(Ahrefs) top pages clicks, positions, keywords, referring domains
prioritize_fixes ranked by traffic impact
1. /pricing is noindex but gets 900 visits a month. Fix this first.
2. /results ranks #6 for "league table" (5,600 searches a month).
Add the term to the H1 to push into the top 3, ~330 extra clicks.
3. /guide shows up 20,000 times at position 3 but gets 0.2% clicks.
Rewrite the title and description, test them with check_snippet.
How it works: an MCP server cannot read other connectors, and should not. OnPage.dev's results and instructions ask the assistant to fetch the numbers from any SEO data tool it has, and the assistant passes them to prioritize_fixes. The numbers are used for that answer only and never stored. OnPage.dev never logs in to those services.
What it looks for: blockers on pages with traffic, low click-through for the position, keywords in striking distance (position 4 to 15), top queries missing from the title or H1, and pages with strong backlinks that score badly.
What makes OnPage.dev different from every other SEO MCP server:
prioritize_fixes ranks every fix by traffic impact with their numbers. How it works.meta-missing or h1-missing, and the main tools declare an output schema. An agent can work through issues one by one and prove each one is gone.scan_html, validate_schema and compare_html test new HTML and JSON-LD before it ships, and catch regressions such as a stray noindex or a removed canonical.check_urls follows every redirect chain for up to 20 URLs at once; test_robots answers "may this bot fetch this URL" with Google's own matching rules, including * and $ wildcards.start_site_audit runs as a job the agent works through, so a 25-page audit fits a free hosted service. It reports issues across the site by code, broken and orphan pages and duplicates.compare_pages puts a page next to up to 3 competitors with content gaps and fixes to catch up; suggest_internal_links shows which pages should link where.onpage://rules resource gives the assistant every check, threshold and fix, so it explains SEO from the same rules the scanner uses.The server is hosted. There is nothing to install or run, and no key.
Server URL: https://onpage.dev/mcp
claude mcp add --transport http onpage https://onpage.dev/mcp
Or install the plugin for the server plus ready commands and skills.
Settings → Connectors → Add custom connector → paste https://onpage.dev/mcp → save. Turn the connector on in a chat.
Add a custom connector with https://onpage.dev/mcp (Settings → Connectors, with developer mode on). Which plans can use custom connectors is up to OpenAI.
~/.cursor/mcp.json or .cursor/mcp.json in your project:
{
"mcpServers": {
"onpage": { "url": "https://onpage.dev/mcp" }
}
}
.vscode/mcp.json:
{
"servers": {
"onpage": { "type": "http", "url": "https://onpage.dev/mcp" }
}
}
Streamable HTTP transport at https://onpage.dev/mcp. JSON responses, no authentication.
21 tools in five jobs. Every result links to the full visual report on onpage.dev. Issues carry stable codes; the full list is in the onpage://rules resource.
| Tool | What it does | Example result |
|---|---|---|
scan_page | Score from 0 to 100, the issues to fix first with why and how, AI readiness and key facts | 46/100, 5 fixes |
deep_audit | Everything measured, by section: speed hints, links and anchor texts, accessibility, image SEO, rich results, security headers, content and technical facts | 8 sections |
get_fix_pack | Ready-to-paste HTML for every issue, written from the page's own content | <title>, <meta>, canonical, social tags |
generate_schema | Article, Product, FAQ, Organization or breadcrumb JSON-LD, validated against Google's rich result rules. Values it cannot read are marked TODO, never invented | Product: eligible |
rescan_and_compare | Scan again after a fix: score change, what got fixed, what is new | 46 → 94 |
| Tool | What it does | Example result |
|---|---|---|
start_site_audit | Audit up to 25 pages from the sitemap as a job. Returns an audit_id and the first progress | 3 of 25 scanned |
get_site_audit | Continue until complete, then read the site-wide results: issues by code with affected pages, broken, orphan and duplicate pages | 25 pages, average 92 |
suggest_internal_links | Which audited pages should link to a page, with anchor text. Works for a new page by topic | 3 links to add |
prioritize_fixes | Rank fixes by traffic impact, using numbers from your connected Search Console, GA4, Ahrefs or Semrush: blockers, low CTR, striking distance, missing keywords, strong links on a weak page. Falls back to severity | #6 → top 3, ~336 clicks/mo |
compare_pages | A page next to up to 3 competitors: side by side, content gaps, structured data they have and fixes to catch up. Flags cookie walls | You lead on 12 of 14 |
| Tool | What it does | Example result |
|---|---|---|
check_ai_visibility | Which AI crawlers may read the page (GPTBot, ClaudeBot, PerplexityBot, Google-Extended and more), llms.txt, 11 checks for AI answers, and the page as a model reads it | 7 of 8 crawlers allowed |
find_answer_passages | Does the page answer a question well enough for AI to quote it? Returns the best passages with length and fit | Best passage, 64 words |
ai_crawler_policy | robots.txt rules for 13 AI crawlers from a policy (allow all, AI search only, block all), merged into your existing file | 13 bots, merged |
generate_llms_txt | A ready llms.txt built from the sitemap and home page | llms.txt, 42 pages |
| Tool | What it does | Example result |
|---|---|---|
check_urls | Status codes and full redirect chains for up to 20 URLs, checked against where each should land. Flags chains and temporary redirects | 19 of 20 OK |
test_robots | May Googlebot, GPTBot, ClaudeBot or any crawler fetch this URL? Uses Google's matching rules and returns the exact deciding rule. Can test a proposed robots.txt too | Blocked by "Disallow: /search?" |
| Tool | What it does | Example result |
|---|---|---|
scan_html | Scan HTML that is not live yet: a local build, a template or a draft. Same score and fixes as scan_page | 92/100 before deploy |
compare_html | SEO diff between two versions of a page: score, issues fixed and introduced (by code), and changes to title, description, H1, canonical, robots and structured data | 72 → 18, noindex added |
validate_schema | Check JSON-LD you wrote: syntax, @context and @type, ISO dates, absolute URLs, placeholders and Google's required fields | 2 errors, 3 warnings |
check_focus_keyword | Eight checks for one search term: title, description, H1, URL, opening text, subheadings, alt texts and keyword use | 6 of 8 |
check_snippet | Pixel-width estimate of title and description against Google's cut-off, desktop and mobile | 612 px, cut off |
One click in clients that show prompts:
onpage://rules: every check with its stable code, why it matters and how to fix it, plus the key thresholds (title and description length in pixels, word counts, alt text, server response, AI answer length).
In clients that support MCP Apps (Claude) or the Apps SDK (ChatGPT), scan results also show as a visual card:
| OnPage.dev | Data vendor MCPs (Ahrefs, Semrush and similar) | Self-hosted audit servers | |
|---|---|---|---|
| Price | Free | Paid plan or API credits | Free |
| Setup | Paste one URL | Account and key | Install and run yourself, often extra API keys |
| Built for | Acting on results: fix, check before deploy, confirm | Research data: keywords, backlinks, rankings | Audits and reports |
| Scan unpublished HTML | Yes, scan_html and compare_html | No | Rarely |
| Agent-grade output | Stable issue codes, output schemas, rules resource | Data rows | Varies |
| Migration checks | Redirect chains, robots.txt tester | Partly | Partly |
| Whole-site audit | Up to 25 pages, free, as a job | Yes, paid | Yes, self-run |
| Competitor gaps | Page vs page, with content gaps | Keyword and backlink gaps | Rarely |
| AI search tools | Crawler access, AI-answer checks, robots.txt policy, llms.txt | AI visibility tracking on some | Rarely |
| Ready code | Fix pack, JSON-LD, robots.txt, llms.txt | No | Varies |
| Uses your traffic data | Yes, from GSC, GA4, Ahrefs or Semrush already in the chat | Their own data | Rarely |
What OnPage.dev does not have: its own search volumes, rankings or backlinks. Connect a data vendor in the same chat and OnPage.dev turns their numbers into ranked fixes.
Why does ChatGPT never mention example.com? Check it with OnPage.dev.
Scan our pricing page and give me the five fixes with the code.
Let ChatGPT search read our site but block AI training. Write the robots.txt.
We moved to a new URL structure. Check that these 20 old URLs redirect to the right new pages.
Review this pull request for SEO regressions: compare the old and new HTML of the changed pages.
Audit our whole site and tell me the three fixes that help the most pages at once.
We are writing a page about trail running shoes. Which of our pages should link to it?
How does our pricing page compare with these two competitors, and what are we missing?
Does our blog post answer "how long does shipping to Germany take" well enough for AI to quote?
Write three title options for this page and check which ones fit in Google.
Fix the SEO issues in this project, check the build with OnPage.dev and keep going until it scores 95.
The server plus commands and skills that run the fix loop for you.
/plugin marketplace add seoonpage/onpage-tools
/plugin install onpage@onpage-dev
/seo-check [url or path]: scan a live page, or find this project's built HTML and scan that, then fix what is wrong./ai-visibility: check whether ChatGPT, Claude, Perplexity and Gemini can read and quote the site, and write robots.txt and llms.txt.metadata, Astro frontmatter), checks with scan_html before deploy and confirms with rescan_and_compare after.About 300 tokens of always-on context.
Scans the HTML your build produces on every pull request, comments the scores and top fixes, and fails when a page drops below your minimum.
name: SEO gate
on: pull_request
permissions:
contents: read
pull-requests: write
jobs:
seo:
runs-on: ubuntu-latest
steps:
- uses: actions/checkout@v5
- run: npm ci && npm run build
- uses: seoonpage/onpage-tools@v1
with:
paths: "dist/**/*.html" # built HTML to check
base-url: "https://example.com" # optional, for canonical and relative links
min-score: "80" # fail below this
max-files: "10" # about 20 scans per minute
The pull request gets a comment like this, updated on every push:
❌ OnPage.dev SEO gate: lowest score 36 (minimum 80)
Page Score AI readiness Fix first dist/about/index.html36 0 There is no meta description; There is no H1; The title is 5 characters dist/index.html90 22 Thin content: 32 words; No structured data; Open Graph is incomplete
scan_html is analysed in memory and dropped. For change tracking, only the score and issue names are kept, under a hashed key, for up to 30 days.prioritize_fixes are used for that answer only. OnPage.dev never connects to Search Console, GA4, Ahrefs or Semrush itself.Full details: onpage.dev/privacy and onpage.dev/terms.
scan_html skips them.check_snippet are an estimate; Google can also rewrite snippets.MIT
This listing does not have a supported local package template. Use the maintainer’s documentation for its hosted endpoint, authentication, and client-specific setup. No install command has been inferred.