The integration problem
Look at the integrations page of any AI tool on the market. You'll see a grid of logos — WhatsApp, Gmail, Google Calendar, Slack, Stripe, Instagram, a dozen others. Each logo is a one-off connection somebody's engineering team built by hand. Each one breaks when the underlying platform changes its API. Each one is a ticket in a backlog somewhere waiting for an engineer to patch it.
We didn't want to build Sami that way. Every "integration" we add by hand becomes tech debt we'll carry for years. And every vertical we want to serve — real estate today, yoga studios tomorrow, driving schools the week after — would force us to rebuild half the stack if the plumbing were that brittle.
So we built Sami on MCP instead.
What MCP actually is
MCP stands for Model Context Protocol. It's an open standard released by Anthropic in late 2024. The one-sentence version: MCP gives AI models a shared way to talk to external tools, so you build the connection once and any MCP-compatible AI can use it.
Think of it like USB. Before USB, every peripheral had its own port — serial, parallel, PS/2, ADB. After USB, you plugged anything into anything. MCP is doing the same thing for AI and tools. One standard. Any tool. Any model.
In MCP terms, each tool is a "server." The AI that calls those tools is a "host." Host asks, server answers. That's the whole architecture.
How Sami is wired
Sami is a host. Everything else is a server.
Your WhatsApp Business account? An MCP server. Your Gmail inbox? An MCP server. Your Google Calendar, your booking system, your property management tool, your CRM — all MCP servers. When a message comes in on WhatsApp, Sami doesn't have any WhatsApp-specific code running inside it. There's a WhatsApp MCP server that hands the message to Sami and sends Sami's reply back out.
The reasoning — the part that decides what to say, when to book, when to escalate — is separated from the channel. WhatsApp, email, SMS, Instagram DMs, voice calls: to Sami, they're all just inputs and outputs. Same brain. Different pipes.
A concrete example
A guest messages a real estate host at 11pm asking if the apartment has a washing machine.
- 01 The WhatsApp MCP server receives the message and forwards it
- 02 Sami figures out what the guest is actually asking
- 03 Sami calls the property MCP server to check the amenities list
- 04 Sami drafts a reply in the host's voice
- 05 The WhatsApp MCP server sends it back
Swap WhatsApp for email at step 1 and step 5. Everything in the middle is identical.
A real example: thereach.ai
thereach.ai, our real estate product, is live today. Under the hood, it's not a monolithic app. It's three things plugged together:
- — Sami host — the agent, the memory, the reasoning
- — Real estate MCP server — property details, amenities, check-in logic, local recommendations
- — Meta and Telnyx MCP servers — the channels: WhatsApp, SMS, and voice
That's the whole system. When we want to improve how Sami handles late-night booking questions, we change Sami. When we want to add a new piece of property data, we change the real estate MCP server. When Meta pushes an update to the WhatsApp Business API, we change the Meta MCP server. Nothing is tangled together.
That separation is why thereach.ai is our proving ground — everything we learn about the host layer there goes straight into every other vertical we build next.
What this unlocks
Here's the practical consequence. If you're a real estate host on thereach.ai and you also run a small yoga studio on the side, we don't need to build you a second product. We write a yoga-studio MCP server — class schedules, pricing, waitlists, drop-in rules — and plug it into the same Sami. Same voice, same memory, same assistant. New domain knowledge. In practice, that's a two-week build, not a six-month rebuild.
Same logic for every vertical we want to serve. Local restaurants? A restaurant MCP server with menus, opening hours, reservation logic. Coaches and therapists? A scheduling and intake MCP server. Driving schools, salons, lawyers — each one becomes a template and an MCP server, not a new app you have to learn.
It also works the other direction. If you already use a tool we haven't integrated — a niche property management system, a local booking platform, your own custom spreadsheet — someone can write an MCP server for it in a weekend, and Sami can use it the following Monday.
The agent stays the same, and the world it can reach keeps growing. The plumbing is boring and standard. The intelligence is the part we focus on.