MCP Integration for Claude, Cursor and ChatGPT

You open Claude Desktop. You ask: "How did my AI visibility move this week, and which sources should I pitch?". Claude pulls the data from Mentionable through MCP, looks at your tracked prompts, sees which domains cite your top two competitors but not you, ranks them by impact score and marketplace price, and gives you a four-line outreach shortlist.

No tab switching. No copy-pasting dashboard screenshots into a chat. No "let me check that in Mentionable and come back". Your AI has the data, because the data goes where AI lives.

What is MCP and why is it in a GEO product

The Model Context Protocol is an open standard introduced by Anthropic in late 2024 that lets AI agents call external tools and data through a common interface. A client (Claude Desktop, Cursor, Claude Code, ChatGPT with custom connectors, Zed) implements the client side. A vendor (Mentionable, in this case) implements a server that exposes tools and resources. The agent discovers the tools, invokes them when relevant to the user's request, reads the response, and incorporates the data into its answer.

For an AI visibility tracker, this is not a nice-to-have. It is the natural distribution channel.

Most marketing tools built in the last decade assumed the dashboard was the product. You pay, you log in, you read charts, you export CSVs. The model made sense when marketers read reports in the morning and took decisions in a UI.

The model does not make sense when the marketer's daily tool is an AI agent. A consultant drafting a blog post inside Cursor, an agency operator using Claude Desktop to run 5 client projects, a founder asking ChatGPT to summarize their weekly metrics: they do not want to leave their AI to check a dashboard. They want the dashboard's answer inside the AI.

MCP is the protocol that makes this work. Mentionable's server makes it real.

The 62 tools, and what they unlock

The MCP server exposes 62 tools today (40 reads, 22 writes), aligned with the highest-signal data Mentionable already tracks. The eight original tools are detailed below; the fifty-four added since launch (scans and raw results, local places, tracked links, AI crawlers, Reddit signals, Page Audit, fan-out extractions, bulk audits, GEO Audit, Recommendations, Action Log, e-commerce retailers and products) are summarized at the end.

list_projects

Your workspace's projects, with tenant scoping applied automatically. The entry point for any agent workflow. Every other tool takes a projectId, so this is the first call the agent makes when you ask "what is happening across my brands".

list_prompts

Every tracked prompt for a project, with its full stats: total runs, mention rate, visibility per LLM (ChatGPT, Claude, Perplexity, Gemini, Copilot, Google AI Mode, Google AI Overview), latest result per LLM with brand position and sentiment. An agent can answer "which prompts am I losing visibility on this month" in one call.

list_llm_sources

The most valuable tool for content strategy. Returns every domain that appeared in LLM responses for your tracked prompts, broken down into three types of appearance:

• Visible citations: the LLM cited this domain in its answer, the user saw the link
• Hidden citations: the LLM read this domain in its context but did not cite it in the final answer (available on LLMs that expose this signal, typically Perplexity and Gemini)
• Fan-out searches: the LLM ran a background search and found this domain in the results, whether it cited it or not

The fan-out data is a GEO goldmine. It tells you which domains the LLM considers when answering prompts in your niche, even if no citation was ever shown to the user. Those domains are the real influence map. An agent can ask the tool for domains with high fan-out count but low visible-citation count, and you immediately see the sites LLMs consult but do not credit: prime candidates for content collaboration or backlink outreach.

list_fan_outs

The queries themselves, not the domains. Every web search the LLMs ran behind the scenes to answer your tracked prompts, deduplicated across runs and ranked by frequency. Each row includes the query text, how many times it was issued, which LLMs issued it, the originating prompts, and first/last seen dates.

No clustering, no heuristic filtering, no "minimum topic size" gate: the raw surface. An agency running a backlink strategy can ask "show me the top 100 fan-out queries on my client's project this quarter, filter to Perplexity only" and get a copy-pasteable list back. Supports search, prompt filter, LLM filter, and two sort modes (frequency, recent).

list_backlink_opportunities

Domains where placing a backlink could lift your GEO visibility, ranked by a composite impact score that weights citation count, fan-out frequency, number of distinct LLMs that cite the domain, and number of your tracked prompts where the domain appears. Each domain comes with its current marketplace offers (price, provider) from Mentionable's aggregated backlink data. An agent can answer "what is my best backlink opportunity under €500" with a single tool call and return a sorted shortlist, not a wall of noise.

list_competitors

The competitors tracked on a project, with aggregated mention totals, Share of Voice, per-LLM presence, and status (confirmed, suggested, rejected). The foundation for competitive intelligence conversations with your AI.

list_competitor_sources

The single most useful outreach endpoint. Given a specific competitor, the tool returns every domain where that competitor gets cited, with mention count, LLMs, top URLs, and sample context from the LLM response. If your competitor is cited 12 times on a specific review site and you are not, that site is a direct outreach target. An agent can prepare a full competitive outreach list in one prompt.

bulk_update_competitor_status

The write tool. Accepts up to 50 competitor updates in a single call, each either CONFIRMED, SUGGESTED (pending) or REJECTED. An agent can run the common operational workflow entirely from chat: "list the 12 competitor suggestions detected last week, keep these 8 as real competitors, reject these 4 as noise". One call, atomic results, per-item error handling when an ID is invalid.

Added since launch: scans, results, local, links, AI crawlers, Reddit, audits, recommendations, Action Log, e-commerce

Fifty-four more tools shipped after the initial launch, covering these adjacent surfaces:

• Scans and raw results. list_scans, list_prompt_results, get_prompt_result. Walk the tracking history scan by scan, list one prompt's answers per LLM and per date, then pull the full payload of a specific result: cited sources (visible and hidden), fan-out searches, competitor mentions, and the raw LLM response (opt-in, because it is the heaviest field).
• Local visibility and tracked links. list_local_places, list_tracked_links. See which businesses ChatGPT and Google AI Overview show in their map for your prompts, and how often the links you bought are cited as sources.
• AI crawlers. get_ai_crawler_overview, list_ai_crawler_pages, get_ai_crawler_page, list_ai_crawler_visits. Once the project is connected to Cloudflare, see which AI crawlers visit your site and which pages they read, split into answer reads, search indexing and training crawls, crossed with the citations of each page in AI answers. For a page published with Mentionable, follow the delay from publication to the first AI visit and the first citation. Visits come from Cloudflare AI Crawl Control, synced every night.
• Reddit signals. list_reddit_threads, get_reddit_thread, bulk_update_reddit_thread_status, enrich_reddit_thread. Surface Reddit threads cited by LLMs in your tracked prompts, triage them at scale, and trigger a Bright Data scrape (50 credits) on the most promising ones. Outreach pipeline in chat instead of in a spreadsheet.
• Page Audit. list_page_audits, run_page_audit, get_page_audit. Queue a Page Audit on any of your pages to get a structured list of GEO fixes, then track and fetch the results. run_page_audit charges 100 credits per audit on every plan and respects the credit cap (capped by monthlyOverageCapEuros). quota_exceeded is returned as structured data so the agent can explain the limit and the upgrade path. Audits run async; poll get_page_audit with the auditId until COMPLETED.
• Fan-out extractions. create_fan_out_extraction, list_fan_out_extractions, get_fan_out_extraction. ChatGPT no longer shows its search queries to logged-out visitors, so an extraction queries its search engine through the OpenAI API for the prompts you pick: queries, ranked results, opened and cited pages, plus a summary of the queries where your site is absent. 5 credits per prompt, charged only when the engine searched; a dry run returns the cost first.
• Bulk audits. list_bulk_audits, get_bulk_audit, list_bulk_audit_competitors, select_bulk_audit_competitors. Follow a bulk audit campaign and curate its competitor shortlist from chat.
• GEO Audit (white-label). list_geo_audits, get_geo_audit, list_geo_audit_competitors, select_geo_audit_competitors. Drive a one-shot prospect audit from chat: read its status, curate the competitor shortlist while it sits in REVIEWING, then freeze the selection to trigger insight generation.
• GEO Audit report editing. list_geo_audit_recommendations, add_geo_audit_recommendation, remove_geo_audit_recommendation, set_geo_audit_recommendation_visibility, list_geo_audit_content_ideas, add_geo_audit_content_idea, remove_geo_audit_content_idea, set_geo_audit_content_idea_visibility, set_geo_audit_section_visibility, list_audit_ctas. Add your own recommendations and content ideas on top of the generated ones, hide the ones that do not apply, and tease a section behind a CTA. Everything the report UI does by hand, callable by an agent.
• Recommendations. list_recommendations, get_recommendation, set_recommendation_status. The actions that would get the project cited, ranked by impact: pages AI engines rely on to recommend your competitors (directory to join, ranking to ask into, Reddit thread to answer), then a step-by-step plan per prompt or local zone, starting with your own page. Your agent works through the plan and marks each step done, so its effect is measured from that day.
• Action Log. list_action_logs, create_action_log. The project journal: every optimization performed (page rewrites, published content, outreach), with dates. Your agent logs what it just did with create_action_log, then list_action_logs filtered on a date range turns a monthly report from a memory exercise into one tool call. Logged actions also show as markers on the dashboard performance chart, so visibility movements line up with the work that caused them.
• Prompts and project settings. create_prompts, update_project_settings. Add prompts to track from chat, and turn a project's e-commerce mode on or off.
• E-commerce: retailers and products. list_retailers, get_retailer, list_products, get_product, get_product_share_of_voice, list_product_prompt_matrix, bulk_update_product_status, update_product, merge_products, split_product. On an e-commerce project, see where AI engines send buyers, which of your products they recommend, and your share of voice by product type against competitors' products. The prompt matrix lists the questions where a competitor's product is cited and none of yours. Clean up products from chat: confirm, rename, merge, split, or change a product's type.

Full schemas for all 62 tools at /docs/mcp/tools.

Real workflows, not demos

The value of MCP shows up in specific workflows that collapse multi-step manual work into a single conversation.

Content briefs from AI citation data

You ask Claude: "Write a content brief for the prompt 'best gym management software'. Use the top 5 cited sources as reference material and explain why each ranks."

Claude calls list_llm_sources filtered on that prompt, gets the top 5 domains with their fan-out queries and sample URLs, reads the titles and seenAs classification, and drafts a brief that structurally mirrors what the LLMs already value. Your brief is not speculation. It is derived from real citation signals.

Outreach lists without spreadsheets

You ask ChatGPT: "Find me 15 outreach targets for this month. I want domains that cite at least two of my three top competitors but have never cited us."

ChatGPT calls list_competitor_sources on each of your three top competitors, intersects the domain lists, compares against list_llm_sources for your own citation footprint, removes domains where you already appear, and returns a ranked list with the competitor citation counts and your current absence score. A manual version of this workflow takes 90 minutes of spreadsheet work.

Operational triage in one conversation

You ask Claude Desktop: "Show me the competitor suggestions detected in the last 14 days. I will review each and confirm or reject."

Claude calls list_competitors filtered on SUGGESTED status recently created, presents them with context and mention count, you approve or reject each from chat, Claude calls bulk_update_competitor_status once with the full batch. Triage that used to eat 20 minutes of clicking in the app, done in a two-minute conversation.

Security model you can reason about

A single principle: an API key can never grant access beyond what the user already has.

Keys are scoped to your TenantMember (your user-tenant pair). If you belong to two workspaces, each workspace has its own set of your keys. You create a key in Settings → MCP, give it a name, optionally restrict it to specific projects within your scope, copy the plaintext once (shown a single time), and configure your AI client.

On every call, the server resolves the effective permissions by intersecting your current membership permissions with the key's optional restrictions. If you later get removed from the workspace, your keys stop working immediately on the next call. If your role gets downgraded to read-only (customer), your keys immediately start rejecting write operations. Nothing is cached, no stale permission grants.

Plaintext is never stored. The server stores a SHA-256 hash of the key and a visible prefix for identification in the UI. Revocation is one click and takes effect on the next call.

Rate limits are applied per key using a sliding window: 100 requests per minute by default, generous enough for normal AI agent traffic.

Setup in under 60 seconds

Claude, ChatGPT, Claude Code and Cursor connect with OAuth. Add a custom connector with the server URL, click Connect, approve on the Mentionable screen:

https://mentionable.ai/api/mcp

There is no key to create or paste. For Claude Code, register the server with claude mcp add --transport http mentionable https://mentionable.ai/api/mcp, then run /mcp and choose Authenticate.

Automations and clients without OAuth

For n8n, scripts or a client without a Connect button, create an API key in Settings → MCP and send it as a Bearer token:

{
  "mcpServers": {
    "mentionable": {
      "url": "https://mentionable.ai/api/mcp",
      "headers": {
        "Authorization": "Bearer mnt_sk_YOUR_KEY"
      }
    }
  }
}

Connectors that do not support custom headers

If your AI connector only accepts a URL (no custom headers), pass the API key as a query parameter:

https://mentionable.ai/api/mcp?key=mnt_sk_YOUR_KEY

Both ?key= and ?apiKey= are accepted. Prefer the Bearer header when your client supports it: query parameters can be logged by servers, CDNs, reverse proxies and browser history, which increases the risk of the key leaking. Rotate the key in Settings → MCP if you suspect exposure.

What makes the data worth querying

An MCP server is only as useful as the data behind it. Building one correctly means committing to a stable tool contract, proper per-user authentication, membership-aware permissions, and a data model worth querying.

Mentionable has had a proper multi-tenant data model and auth layer from day one, because the app itself is multi-user. Adding MCP was a natural extension, not a retrofit. What your AI gets is daily tracking across 7 AI engines, the fan-out queries behind each answer, visible and hidden sources, competitors and audits, in one consistent contract.

The bigger bet

The tools that survive the LLM transition are the ones that stop assuming the human will open their UI. The product is the data. The UI is one interface. MCP is another. Tomorrow there will be more.

If you run marketing in 2026 and your AI assistant cannot access your visibility data without you copy-pasting, you are paying for the wrong tool.

AI assistant or MCP?

You do not need MCP to ask questions about your data. The Mentionable AI assistant works inside the dashboard with no setup, on every plan: it reads your project, explains the numbers and links to the page where you act. MCP brings the same kind of data into your own AI, on the Pro and Agency plans, when you want your agent to use it in briefs, reports or automations next to the rest of your stack.

Try it

Pick the Pro or Agency plan, add https://mentionable.ai/api/mcp as a connector in Claude or ChatGPT, approve the connection, and ask your AI your first real question about your visibility. On the free trial, start with the AI assistant in the dashboard.

Related articles

• Competitor Tracking - the underlying tracking that makes the MCP competitor tools meaningful.
• Source & Citation Tracking - the data behind list_llm_sources.
• Outreach Opportunities - the in-app version of the backlink workflow that the MCP also exposes.
• AI Assistant - ask questions about the same data inside the dashboard, on every plan.