Skip to content
← All casesIntegration HubAI

92 Integrations, One Terminal

MCP infrastructure for AI agents: 11 servers, 6 LLM APIs, 62 Google APIs — agents work with any external service

Problem

What doesn't work

30+ daily switches between tools: Telegram, email, servers, clouds, Figma, GitHub. Each one — open tab, log in, find what you need. 5 min × 30 times = 2.5 hours of lost time.

Solution

Architectural approach

Single entry point via Claude Code CLI with MCP servers. 11 MCP connections (4 local + 7 cloud), 6 LLM APIs, 62 Google APIs, 4 servers, 4 email channels. One terminal command instead of 5 minutes of switching.

Challenges

What made it hard

OAuth2 authorization for each service — a separate pain: tokens expire, scopes differ, refresh mechanisms are non-standard. Telegram User API (Telethon) requires 2FA on every new session — had to build a re-auth pipeline. 92 integrations mean 92 potential failure points — monitoring and graceful fallback for each.

Role

My role & contribution

Architect & developer

Designed MCP infrastructure, wrote custom MCP servers (Telegram User API on Telethon, Google Drive, Yandex Disk/Mail on FastMCP). Configured OAuth2 for cloud MCPs (Figma, Canva, Gmail, Calendar). Integrated 6 LLM APIs, 4 servers (SSH/paramiko), 62 Google APIs across 2 GCP projects.

Demo

How it looks

Screenshots

Real screenshots

Architecture

System architecture

Integration Hub ArchitectureClaude Code CLIorchestration layerMCP SERVERS (LOCAL)CLOUD MCPLLM APIsSERVERS (SSH)TelegramGoogle DriveYandex DiskYandex MailFigmaCanvaGmailCalendarOpenAIAnthropicGoogleDeepSeekLangFuseLiteLLMmars95.81.xbanki62.109.xbanki2176.123.xdeploySSH key92 integrationsAI / LLMDataInfraEval
Implementation

How it works

Local MCP servers (Python + FastMCP) for Telegram User API, Google Drive, Yandex Disk, Yandex Mail. Cloud MCP via claude.ai (Figma, Canva, Gmail, Calendar). Cross-session memory via file system. SSH deploy via paramiko.

Architecture Decision

Why this way

MCP servers instead of direct API calls

Alternative

Write wrapper scripts for each service and call via bash

Why it didn't fit

Scripts: each is a separate context, no discovery, no typing. MCP: unified protocol, Claude sees all tools, can combine them.

Result

Single entry point. Claude picks the right tool for the task

Metrics

Results

01
Context switch: 5 min → 5 sec
02
Save ~2.5 hours daily
03
92 integrations through one entry point
04
11 MCP servers + 6 LLM APIs
05
62 Google APIs across 2 GCP projects
Business Impact

Impact on business

At 30+ switches/day, saves ~2.5 hours daily. Monthly — ~50 hours or 6+ workdays. One CLI command replaces: open browser → login → navigate → search → act.

Methods

Algorithms & patterns

MCP ProtocolOAuth2Telethon (User API)FastMCPSSH/paramiko
Stack

Technologies

  • Claude Code
  • MCP Protocol
  • Python
  • FastMCP
  • Telethon
  • OAuth2

Ready to discuss?

If you need an architect who builds autonomous AI systems — reach out.

Serbia-based · CET/CEST timezone · EU-aligned working hours · International contracts experience