Claude Desktop (MCP) Setup

Connect Cited to Claude Desktop with MCP

Model Context Protocol (MCP) is an open standard that lets AI assistants like Claude communicate directly with external tools and data sources. Instead of copying and pasting information between applications, MCP creates a live connection so Claude can pull data from Cited, run audits, and deliver recommendations inside your conversation.

With the Cited MCP integration, Claude Desktop becomes your GEO copilot. Ask questions in plain English and get actionable insights about your AI search visibility without ever leaving the chat window.

Setup Options

There are two ways to connect Cited to Claude Desktop. Choose the option that fits your workflow.

Option A: Claude Desktop Custom Connector (Recommended)

This is the fastest way to get started. No command-line tools required.

1. Open Claude Desktop and go to Settings > Customize > Connectors.
2. Click Add Connector.
3. Enter the following URL:

https://mcp.youcited.com/mcp

4. Save the connector. Claude Desktop will handle the rest.

This method uses Anthropic's built-in connector system and is the simplest to maintain. Updates to the Cited MCP server are picked up automatically.

Option B: Remote via npx

If you prefer to configure MCP servers through the Claude Desktop config file, you can use the mcp-remote package to connect to the hosted Cited server.

Add the following to your Claude Desktop MCP configuration:

{
  "mcpServers": {
    "cited": {
      "command": "npx",
      "args": ["mcp-remote", "https://mcp.youcited.com/mcp"]
    }
  }
}

This approach requires Node.js to be installed on your machine. The mcp-remote package is fetched automatically by npx the first time you connect.

Authentication

When you first use a Cited tool in Claude Desktop, an OAuth flow opens your default browser and prompts you to log in to your Cited account. After you authorize access, the session token is stored locally so you do not need to log in again until the token expires.

Scopes

The Cited MCP server requests a single coarse cited scope at authorization time. Claude Desktop displays this in the consent prompt. Granting it gives Claude access to every tool the MCP server exposes, scoped to your user — Claude cannot see businesses, audits, or recommendations belonging to other users.

If you're inspecting the OAuth metadata directly:

curl -s https://mcp.youcited.com/.well-known/oauth-authorization-server | jq .scopes_supported
# → ["cited"]

Cited's resource metadata (at https://youcited.com/.well-known/oauth-protected-resource) advertises the same single cited scope, in line with RFC 9728 §3.3 (the resource field matches the metadata host). When Cited adds finer-grained scopes (cited:business:read, cited:audit:write, etc.) we'll publish them to both endpoints simultaneously.

Example Workflows

The Cited MCP server exposes 47 tools spanning the full GEO lifecycle — from probing how AI sees your business today, to running audits, to generating recommendations, to closing the loop on actions. Once connected, you can interact with the platform through natural conversation. The examples below walk through a typical user journey: orient → diagnose → execute → measure → close the loop → share. Don't name the tool to Claude — describe what you want, and Claude picks the right tool (or chain of tools) on its own.

Quick health check on your account

"Quick health check on my Cited account."

Claude calls check_auth_status and surfaces a snapshot: account email, plan tier, and current utilization across businesses, audit runs, recommendation jobs, and solution jobs. Proactive flags fire when you're approaching a plan cap (for example, "you're one business away from the cap; you've got room for exactly one more before you'd need to upgrade").

What does Cited know about a business

"What does Cited currently know about my [business]? Give me an overview — what's good and what's missing."

Claude calls get_business_hq for a single comprehensive snapshot — five health scores (brand confidence, crawl coverage, AI readiness, citation readiness, buyer clarity), recent audits, recommendations, competitive position, agentic readiness — and synthesizes a structured what's working / what's missing / most leveraged next move summary. If the data looks stale, Claude offers to chain crawl_business to refresh before re-fetching. Useful for stepping into an account fresh, prepping for a stakeholder review, or auditing your own setup.

Get a full action plan

"What should I do to improve my GEO?"

Claude calls list_priority_actions and returns a prioritized list of actions ranked by impact and effort. Each action includes a description, expected improvement, and implementation guidance.

Note: The Action Plan tools are wired into the in-app GEO Assistant today. Exposure via the Claude Desktop MCP server is rolling out with the next cited-mcp release. Until then, use the in-app chat sidebar for the close-the-loop flow described below.

Find quick wins

"Show me quick wins."

Claude calls list_priority_actions and filters for low-effort, high-impact items so you see the changes that deliver the fastest results first. Pair with get_priority_action_summary for an aggregate view of how many quick wins are outstanding.

Run a complete audit

"Run an audit for my business."

Claude triggers the full audit pipeline: crawling your site, analyzing structured data, evaluating content quality, checking AI crawler access, and benchmarking against competitors. Results are delivered as a structured report with scores and recommendations.

Track week-over-week changes

"What changed in our AI search visibility since the last audit?"

Claude calls list_audits to find your two most recent audits on the same template, then compare_audits to compute the delta. The response leads with regressions (questions where score fell), surfaces wins, and quantifies aggregate KPI movement so you see direction at a glance. When everything zeros out in a single cycle, Claude flags the structural-cause shortlist (crawler access, schema/canonical changes, robots.txt, CDN, DNS) before you assume it's a real competitive loss.

Probe positioning before committing to a full audit

"Would AI recommend us if a fintech CTO searched for 'best subscription billing platform with global multi-currency support'?"

Claude calls buyer_fit_query with the buyer profile and your business, returning an ordered fit recommendation list with the matching strengths and gaps. Faster and cheaper than a full audit, and useful for testing positioning, validating a new buyer persona, or benchmarking against a specific search intent before you invest in a deeper run.

Close the loop on actions

"I've done the schema fix, what's next?"

Claude calls update_priority_action_status to mark the action complete — and for recommendation-sourced actions, the Validation Engine auto-runs to verify the change actually landed (green ✓ verified, amber ⚠ couldn't see it yet, grey ? couldn't run). Claude then calls list_priority_actions to refresh your list with updated priorities. This creates a working loop where you can knock out tasks one at a time with verification at each step.

Share an audit as a PDF

"Share my most recent audit as a PDF I can forward to my team."

Claude calls export_audit and returns a presigned download URL with a one-hour TTL. The link is rendered inline so you can click through directly, the suggested filename is included, and Claude flags the expiry timestamp so you know to download or forward promptly. Optional provider argument scopes the export to a single AI engine's view (OpenAI, Gemini, or Perplexity) when you want a per-provider report.

Cross-prompt continuity

The connector keeps a working memory of the conversation, not just isolated tool calls. If your audit comparison earlier in a chat surfaced an unusual zero-everything result, asking "share that audit as a PDF" later will prompt Claude to add a context note to the share — "this is the audit where every metric came back at zero; you may want to attach a one-line note explaining it looks like a structural collapse, not a real competitive loss, otherwise the headline numbers will read worse than reality." That continuity is where the conversational interface earns its keep over a dashboard.

Available Tool Categories

The Cited MCP server exposes 47 tools organized into eight categories:

• Account & Auth -- ping, check_auth_status, login, logout, whats_new, get_pricing, upgrade_plan. Connect, verify, see plan utilization, detect tool-surface changes mid-conversation.
• Business -- Manage business profiles, domains, settings, crawls. Look up business details, trigger fresh crawls, check health scores.
• Audit -- Run GEO audits across multiple AI platforms (ChatGPT, Perplexity, Gemini, Copilot). Manage reusable audit templates, review past results, drill into per-question detail, export PDFs.
• Recommendations -- Generate AI-powered recommendations from a completed audit, with structured insights (question coverage gaps, head-to-head competitor comparisons, strengthening tips, priority actions). Each item carries impact scores and implementation guidance.
• Solutions -- Generate implementation-ready content, schema markup, and deployment artifacts from a recommendation insight. Run individually or in batches.
• Action Plan -- Work a prioritized task list. Mark actions complete (with auto-validation when the source is a recommendation), dismiss items, refresh priorities, track progress.
• Analytics -- KPI trends, combined dashboards, audit-vs-audit comparisons. Track citation rates, share of voice, performance benchmarks, and week-over-week movement.
• Business Intelligence & Probes -- Pull structured intel about a business without running a fresh audit (facts, claims, competitive comparison, semantic health diagnostics, full HQ dashboard) and run ad-hoc buyer-fit simulations against custom buyer profiles. The agent-query layer lets Claude answer "what does Cited know about X" and "would AI recommend us for Y" without burning audit budget.

The full tool list is available via the whats_new tool inside Claude (call it with no arguments to see the most recent release notes, or pass since_fingerprint / since_version to diff against an earlier connection).

Troubleshooting

If Claude Desktop does not show Cited tools after setup, verify the following:

• Your Claude Desktop version supports MCP (version 1.0 or later).
• For Option B, Node.js is installed and npx is available in your system PATH.
• Your firewall or VPN allows outbound connections to mcp.youcited.com.

If the OAuth login window does not appear, try restarting Claude Desktop and initiating a new conversation. The authorization prompt is triggered the first time a Cited tool is invoked in a session.

If the OAuth callback completes in your browser but Claude Desktop doesn't pick up the token, the callback URL may have been cached at the old /auth/oauth-complete path. Hard-refresh the callback page; the current path is /oauth-complete (no /auth/ prefix). Clear browser cache for youcited.com if the issue persists.