MCP-Enabled Integration Platform for CRMs & Productivity Tools

A multi-tenant integration layer that connects CRMs (Zoho, Odoo, HubSpot) and productivity tools (Google Sheets, Gmail, Google Calendar) through a single unified API — consumable both as a standard REST service and as MCP tools for AI agents like OpenManus/Manus.

Built collaboratively with a team of 3.

Why

Every CRM and productivity tool has its own API schema, auth flow, and data model (e.g. lead_name in Zoho vs contact.name in HubSpot). This project abstracts that away behind one consistent interface, so a developer — or an AI agent — can call crm.create_lead without knowing or caring which underlying platform handles it.

Related MCP server: 0nMCP

Architecture

• Transport layer: FastAPI REST endpoints + an MCP server layer, so the same platform serves both human API calls and AI agent tool calls.

• Router structure: Modular APIRouter per integration (Zoho, Odoo, HubSpot, Google Sheets, Gmail, Google Calendar), each exposing create / update / delete / list.

• Multi-tenancy: Each user/org is an isolated tenant with its own credentials and config, backed by a SQL database.

• Unified data model: A mapping layer normalizes platform-specific fields into one standard schema before returning data to callers.

• Provisioning: Secure endpoints for API keys, tokens, and client secrets, including OAuth refresh-token flows for Google and HubSpot.

Design inspired by Supaglue and Apache Camel K.

Tech Stack

Python, FastAPI, MCP (Model Context Protocol), SQL, OAuth2

Testing & Validation

• REST endpoints manually tested via FastAPI's Swagger UI (/docs)

• MCP tool discovery and execution verified using MCP Inspector

• End-to-end flows validated across both human-driven and agent-driven calls

What I built

My individual contributions to this platform:

• Google Maps Tool — Integrated the Google Cloud Console API to give Manus real-time geographical query capability, including text search, nearby search, and place photo retrieval. Implemented secure API key management and request handling.

• Gmail MCP Integration — Registered and integrated an open-source Gmail MCP server (Node.js) as a callable tool for Manus:

  • Local (stdio): Ran the server via npx, debugged initialization handshakes, and configured OpenManus to register it for local tool calls.

  • Remote (SSE): Debugged and fixed the full remote transport layer, including:

    • A failing client/server handshake over SSE, isolated and resolved by tracing each layer of the connection independently.

    • A session-ID bug where the client wasn't forwarding the session ID in request headers, breaking the Gmail server's session tracking.

    • A two-step connection sequence (POST to initialize + persistent GET for the SSE stream) that had to be implemented in the correct order to establish the session.

    • A deadlock in tool discovery, where the client blocked waiting on an SSE reader that hadn't been triggered yet — fixed by having the POST writer push the response directly into the stream before invoking the SSE reader.

• HubSpot Connector — Rebuilt the HubSpot integration as a standalone FastAPI service exposed via MCP, containerized it with Docker, and integrated it into the broader platform as one of the pluggable CRM connectors.

Team vs. individual work

The overall platform architecture (multi-tenancy, unified schema, Zoho/Odoo connectors) was a team effort with 2 collaborators. The Google Maps tool, Gmail MCP integration, and the FastAPI + Docker HubSpot connector above were my individual contributions.