Flusia CRM — MCP Server

2. Features and use cases

Exposed resources (212 tools total):

Every tool declares the standard MCP annotations (readOnlyHint for reads, destructiveHint for irreversible operations) as required by the Anthropic MCP Directory Policy.

Typical use cases

• Conversational sales assistant: search leads, update status, create follow-up tasks, send WhatsApp messages.
• Automatic reporting from natural prompt ("Summary of deals won this week by pipeline X").
• Safe bulk operations: the AI proposes, the user confirms (human-in-the-loop on destructive tools).
• Customer onboarding: lead + deal + task + checklist in a single conversational turn.
• Diagnostics: the AI queries CRM state to investigate anomalies reported by operators.

3. How to connect

Claude.ai (web)

1. Go to claude.ai → Settings → Connectors → Add custom connector
2. Name: "Flusia CRM"
3. Server URL: https://api.flusia.it/mcp
4. Confirm. On first use Claude will start the OAuth flow.

Claude Desktop

In ~/Library/Application Support/Claude/claude_desktop_config.json (macOS):

{
  "mcpServers": {
    "flusia-crm": {
      "url": "https://api.flusia.it/mcp"
    }
  }
}

Restart Claude Desktop. On the first tool call the OAuth flow will open in the browser.

Claude Code

claude mcp add flusia-crm https://api.flusia.it/mcp

Static API token (server-to-server)

For server-to-server integrations (scripts, automations) generate a permanent API token:

1. Sign in to app.flusia.it → Settings → API Tokens
2. Create a new token with the required permissions (per resource: read/write/delete)
3. Use the token as Authorization: Bearer flusia_<token>

4. OAuth 2.1

The server exposes the standard discovery metadata:

Features

• Grant types: authorization_code + refresh_token
• PKCE required (S256)
• Token endpoint auth: none (public client)
• Scope: mcp:full (capped by user CRM permissions)
• All valid HTTPS redirect URIs are accepted in production

Pre-configured supported callbacks

• https://claude.ai/api/mcp/auth_callback
• https://claude.com/api/mcp/auth_callback
• http://localhost:6274/oauth/callback (MCP Inspector)
• http://localhost:6274/oauth/callback/debug

5. Usage examples

The AI confirms with the 2 created IDs and a verbal summary.

Revenue closed last month: €127,430
Win rate: 34%
Top 3 deals: ACME Corp €45,000, Studio Verdi €28,500, TechStart €19,800

The user sees the confirmation prompt for each operation, the AI shows the final summary.

Tools with destructiveHint: true (delete, remove, cancel, archive, unassign) automatically trigger the human-in-the-loop pattern in MCP clients, which ask user confirmation before execution.

6. Privacy and data protection

Data processing follows:

• Privacy Policy
• Terms and Conditions
• DPA (Data Processing Agreement) on request at privacy@flusia.it

What flows through MCP

• Only requests/responses to MCP tools explicitly invoked by the AI
• Logs include only organization_id, tool name, timing and status code — not user payloads
• Log retention: 30 days
• Server: OVH Roubaix (France, EU)
• Application data: Supabase EU (Frankfurt, eu-central-1)

What does NOT flow through

• The Flusia server does not receive the user-AI conversation, only the arguments of the tools the AI chooses to invoke
• Anthropic (claude.ai) can see tool responses in the conversation, but has no direct access to the Flusia database

7. Support

SLA: response within 1 business day for standard support, within 4 business hours for confirmed security issues.