One three-server desk stopped asking an agent to build a page and asked "which page the data says is missing." A September 9 list named eight SEO and agent MCP pieces on a different desk. The choice is which servers the job actually needs.
MCP (Model Context Protocol) servers are the plugs that let an assistant such as Claude Code or Cursor pull data itself. The three-server desk and the eight-piece list are two compositions operators published in September.
Your Own Site First: Search Console Plus Analytics
Search Console MCP servers connect Claude Desktop, Cursor, or Gemini CLI to the Search Console account for pages you own. They expose average position, clicks, impressions, and click-through rate, grouped by query, page, country, device, and date, plus URL inspection and indexing status, sitemap handling, and queries across several sites at once. The September 9 list's line for this plug is "Queries, clicks, indexing without CSV paste."
Those servers can pull up to 50,000 rows in one query, past the 1,000-row ceiling of the web interface. Reading queries, impressions, clicks, and position on your own pages is that same feed.
Analytics is what happened after the click. Google's official open-source Analytics MCP server, in the googleanalytics/google-analytics-mcp repository, exposes account and property info plus core reports, so an assistant can ask in plain language about GA4 traffic, behavior, and conversions. GA4 is Google Analytics 4, the current Analytics property for site traffic and conversions. The path in is Claude, Cursor, or ChatGPT, through a desktop config or a no-code router such as Stape, Porter, Windsor, or Composio. The September 9 list's line for this plug is "First-party traffic and behavior for agent audits."
What Each Piece Adds and What It Burns
| Server | Adds | Burns |
|---|---|---|
| Search Console MCP | Queries, clicks, impressions, position, indexing, sitemaps, several sites at once | Your Search Console account. A September 15 connector was described as free. |
| Analytics MCP | Account and property info, core reports on traffic, behavior, conversions | Your Analytics account. Desktop config or a router. The September 9 list names no per-call price. |
| Live-data MCP (DataForSEO or Ahrefs) | Live results, keywords, backlinks, and the AI answer block when Search Console is not enough. Ahrefs also covers audit inside the agent | Pay per use on the DataForSEO side. A September 15 reply put the minimum deposit at $50. The list names no Ahrefs price. |
| Chrome DevTools MCP | Core Web Vitals, network, and console on a live Chrome | No per-call price on the September 9 list. The agent drives a Chrome instance. |
| On-page analyzers and a 28-rule local crawl | Open-source on-page and technical checks; local crawl and audit | Called local on the list. No price named there. |
| Skills catalog | Search, preview, and install workflows into Cursor or Claude Code | No price named on the list. |
The live-data row uses the September 9 label for DataForSEO: "Live SERP, keyword, backlink, and AI Overview data when GSC is not enough." GSC is Search Console. SERP means the live results page for a query, and the AI Overview in that line is the AI answer block on the page. Ahrefs is the other name for the same slot: keywords, backlinks, audit, and results inside the agent. Chrome DevTools on that list "Complements Playwright, does not replace it." The list also names Clarity and SeoMCP alongside the eight.
Core Web Vitals are Google's loading, responsiveness, and stability signals. Chrome's official server, ChromeDevTools/chrome-devtools-mcp, lets a coding agent drive a live Chrome instance through the Chrome DevTools Protocol (CDP). It runs slow-4G and mobile emulation, trace capture, a Largest Contentful Paint breakdown (when the main content appears), Interaction to Next Paint tests (the delay after a click or a keypress), Cumulative Layout Shift monitoring (layout jump), and compressed trace summaries for the agent loop.
The live-data row is the one that bills with use. A September 15 note on self-hosted OpenSEO plus the DataForSEO API describes the bill as pay per usage, set against a large monthly invoice. Someone the same day put DataForSEO's minimum deposit at $50, called it pay-as-you-go, and said the deposit does not expire. A September 22 desk put API credits for its Claude operation at about $100 a month, and tools before that at $500 or more a month. Once that slot bills per call, put a ceiling on the agent before it runs.
qequals has no MCP server. A suggestions call returns the autocomplete phrases for a seed query, and a search call returns the live results for that query. When an MCP server earns the switch is a different decision from which fields the desk needs.
Which Page the Data Says Is Missing
On September 15, three MCP servers were plugged into Claude Code: Google Analytics for what people landed on, PostHog for what they did next, and DataForSEO for what they searched. The operator stopped asking the agent to build a page and asked "which page the data says is missing." The agent also found a broken analytics setup the operator had not asked about. That desk already included the live-data server.
A French internal-linking playbook from the same day starts on Search Console. Through a free connector (Composio) it pulls three months of page and query: impressions, clicks, click-through rate, and position. It joins that to a crawl from Screaming Frog or a DataForSEO crawl server, with URLs, HTTP status, canonicals, indexation directives, depth, internal links, and the main content. The filter keeps indexable HTTP 200 pages with a coherent canonical, relevant queries at an average position of 5 to 20, enough impressions to be an opportunity, few relevant internal source pages or in-body links, and a real business interest.
Embeddings (numeric similarity scores between pages) preselect source pages on a related subject. The playbook recommends OpenAI's text-embedding-3-small for that step and calls it high-performing and very cheap. Closeness is only the shortlist. The agent checks whether the link would help the reader, whether the source already points at the target, and whether the destination matches the anchor. Their example is a guide on preparing a first trail run recommending a hydration-pack category inside the gear passage. Each proposal returns source URL, target URL, original passage, modified passage, anchor, and justification, from a passage actually on the page. The operator can re-read and apply each change, then recrawl and track the pages and queries in Search Console with the date kept. The reason for staying on that feed: the demand is already visible in Search Console, and the work strengthens pages that exist.
Someone weighing pay-per-go DataForSEO against HaloScan said the two are not the same thing.
The 3-Piece Desk and the 8-Piece Desk
The three-server shape that found the missing page was Analytics, PostHog, and DataForSEO. The internal-link shape was Search Console plus a crawl. The eight pieces in the table are the September 9 week-2 list, and its own close is "Still not a founder starter pack. Stuff that survives a real crawl or a real eval."
Four September 16 notes each describe one desk, separate from the week-2 list.
- One stack lists Hermes running on GPT-Sol, DataForSEO, Search Console and the Google Ads API, Firecrawl and Exa, and skills named Content Humanizer, Last30days, and a custom skill for answer-engine optimization and generative-engine optimization.
- A Dify Marketplace announcement lists Similarweb and Ahrefs to track competitors, Semrush and DataForSEO to find search gaps, X and Reddit to read the room, and Apollo to find prospects. A reply the same day calls it one plugin and one key.
- One note describes openSEO with a Search Console connection through Composio, so Grok or Claude can research keywords, competitions, and opportunities.
- One note says Claude SEO connects to Google APIs and DataForSEO.
Write down the job you run every week. If it starts from pages you own, plug in Search Console and Analytics. If it needs live results, keywords, backlinks, or an AI answer block beyond those pages, add one live-data server. Move to the eight-piece list when the job is what that list called a real crawl or a real eval.
