Overview

• Purpose: Help partners and developers build an MCP server that Bonnie can discover and call as external tools to augment voice/chat assistants.
• Who should read: Backend engineers or integrators familiar with HTTP/JSON building services; basic knowledge of JSON Schema is helpful.
• What you’ll build: An MCP server exposing tools via JSON‑RPC with optional UI discovery, scoping by call state/channel, and a minimal config contract.

What is an MCP server and why it matters for Bonnie

A Model Context Protocol (MCP) server is a lightweight service that exposes callable “tools” over JSON‑RPC, discoverable via tools/list and executable via tools/call. In our setup, Bonnie acts as the MCP client: at the start of a call or chat, she lists available tools, applies scope rules, and then invokes the right tool in‑flow. This lets partners cleanly package capabilities—like reservations, CRM actions, payments, delivery lookups, calendars, or custom business logic—behind clear JSON Schemas and predictable responses.

How this expands Bonnie’s phone and WhatsApp skills

By connecting Bonnie to one or more MCP servers, you can safely extend her capabilities across third‑party platforms without changing Bonnie’s core. Tools can be scoped to call state (setup, in‑progress, completed) and channel (phone, whatsapp) so the right actions are available at the right moment—for example, auto‑fetching customer context during setup, checking table availability mid‑call, or sending a WhatsApp confirmation after booking. Results can be structured JSON, enabling Bonnie to reliably summarize outcomes to the caller, update systems (e.g., SevenRooms, Formitable, Zenchef), or trigger follow‑up messages. Authentication and headers carry restaurant- and conversation‑specific context, keeping integrations secure and targeted while remaining simple to adopt.

Concepts

• MCP in brief: Servers expose tools over JSON‑RPC.
  • tools/list: enumerate available tools and their input schema.
  • tools/call: execute a tool with arguments.
• Bonnie’s role: Acts as the MCP client. It discovers tools at runtime and invokes them during conversations.
• Transports supported: streamable_http (default) and sse.
• Authentication: Optional Bearer token via Authorization: Bearer <token>; additional static headers are also supported.

Architecture and end‑to‑end flow

1. Assistant configuration registers one or more MCP servers.
2. On startup of an interaction, Bonnie lists tools from each configured server and normalizes them to OpenAI‑style function tools.
3. Availability is computed by merging server‑level defaults (from assistant config) with per‑tool meta scoping (from your server).
4. At runtime, Bonnie selects eligible tools for the current call state/channel and calls them via tools/call.
5. Responses can be plain text or structured data; structured data is preferred when available.

Building a server (requirements)