MCP Hub: Model Context Protocol Server for GEO/AEO Agencies | AI Labs Audit

An MCP (Model Context Protocol) server exposes a set of tools and resources that an AI client can consume (Claude Desktop, Cursor, Windsurf, Continue, etc.). AI Labs Audit's MCP server ships 226 tools covering the entire GEO/AEO lifecycle: client creation, audit launching, custom prompts, multi-judge scoring, PDF export, credit management. You drive the platform from your favourite AI instead of opening a dashboard.

Step 1: create an MCP API key from your AI Labs Audit dashboard (API Keys page). Step 2: edit Claude Desktop's config file (claude_desktop_config.json) and add the AI Labs Audit MCP server entry with your key. Step 3: restart Claude Desktop, open a new conversation, and you'll see the 226 tools in the picker. The step-by-step tutorial covers Claude, Cursor and Windsurf.

Agency keys grant access to all 226 tools: user management, credit pool allocation, client create/delete, audit launching, prompt configuration. End-client keys are restricted to read-only operations on their own data: viewing audit results, exporting reports, tracking their scoring. This separation prevents a leaked client-side key from reaching sensitive operations.

Both coexist and cover the same features. The REST API is ideal for integrating AI Labs Audit in a custom backend, a cron, n8n, Make or a CI pipeline. MCP shines for AI-driven natural-language workflows: a consultant types an instruction in Claude Desktop and the AI orchestrates the calls. If you write code, pick REST. If you want chat-driven control, pick MCP.

MCP server access is included from the Agent plan (349 €/month). The free Découverte plan and the Consultant plan (79 €) don't grant access, while Consultant+ (179 €) offers limited read-only tooling. The Agence+ plan (599 €) unlocks all 226 tools, including multi-user credit pool allocation and fine-grained per-member permissions.

MCP is an open standard: any MCP-compatible client can connect. AI Labs Audit is continuously tested with Claude Desktop (Anthropic), Cursor, Windsurf, Continue, ChatGPT Desktop (since its MCP support in late 2025) and several open-source clients. The integration tutorial provides config for the main market clients.

Authentication uses an opaque API key (with a distinct prefix for agency vs client scope). The key is passed in an HTTP header during the MCP handshake. Each key is instantly revocable from the dashboard, and every call is logged (who, when, which tool, which client) in the audit log accessible to the agency admin. Rotating keys every 90 days is recommended.

Limits vary by plan and operation type. Read-only operations (viewing clients, audits, scoring) have a comfortable limit fit for interactive use. Costly write operations (launching audits, generating PDF reports) are gated by the plan's credit quota rather than a call count. On overrun, the server returns a standard MCP code and the client AI can surface the message to the user.

Yes. When you generate an MCP API key, you pick its scope (agency or end-client) and exact perimeter (one client, multiple clients, all resources). An end-client key can only consume read-only tools on its own data. It's the ideal way to give a client an 'AI-drivable' access without risking your agency data or pool credits.

A typical prompt: "Create a client named Acme with the acme.com domain, add the 12 e-commerce library prompts, launch a GEO audit on Claude Sonnet, GPT-4o and Gemini, wait for completion (poll every 30 seconds), then export a PDF report in French. Give me the AGS score, the Hero Grade and the top 3 priority actions." The AI chains the MCP tool calls and returns the synthesis. The tutorial ships 10 battle-tested prompts (one-shot audit, recurring audit, competitor comparison, monthly tracking, etc.).