Zoho CRM MCP demo

A small MCP server exposing four Zoho CRM operations (Deals and Leads modules, real Zoho API v8 field names) as tools an LLM/agent can call in plain language:

• search_deals(stage=None, min_amount=None) — filter the pipeline by stage/amount
• summarize_pipeline() — deal count + total Amount per stage, open vs. closed
• find_duplicate_leads() — groups Leads by (Email, Company) to catch the same contact entering through two different capture channels with slightly different name spelling (e.g. a Marketplace extension form vs. a webhook)
• create_lead(first_name, last_name, company, email, lead_source) — adds a Lead

Why this exists

Built as a short, self-contained proof of concept — not a submission tied to any specific client work. It runs entirely against the synthetic dataset in data/seed.json; there is no connection to a real Zoho CRM org and no CRMOZ or client data involved anywhere.

It mirrors the same pattern as production MCP work already shipped elsewhere (n8n MCP integrations, a GoHighLevel MCP integration for a live client) — applied here to Zoho CRM's schema, to show the pattern fits, not to replace any existing "AI + Zoho" tooling.

Run it

python3 -m venv .venv && source .venv/bin/activate
pip install -r requirements.txt
python3 server.py

This starts the server on stdio, ready to be added to any MCP client (Claude Desktop, Claude Code, etc.) config as a local server.

Verified working

All four tools were exercised directly and through mcp.list_tools() — search_deals, summarize_pipeline, find_duplicate_leads, and create_lead all registered and returned correct output against the seed data, including a deliberately planted duplicate lead (sam.cutter@bramwellfoods.example, entered once via an "Email to Lead" style channel and once via a "Google Sheets" style channel with the name spelled "Sam" vs. "Samuel") that find_duplicate_leads() correctly caught.