MCP (Model Context Protocol) is an open protocol, created by Anthropic in 2024 and adopted in 2025-2026 by OpenAI, Google, and major IDEs, that standardizes how an AI agent discovers and calls external application tools. Applied to WhatsApp Business, MCP allows Claude, ChatGPT/Codex, Cursor, or a custom agent to search chats, send messages, and move contacts in your CRM without glue code—simply point the agent to an MCP server like Aurora Inbox and paste a key.
In one sentence
MCP is to AI agents what USB-C is to devices: a single, standard connector that replaces dozens of custom integrations.
Why was MCP created?
Before MCP, connecting an AI agent to WhatsApp involved one of three paths:
- Function calling owner — define tools in the OpenAI, Anthropic or Google format, one different per provider.
- Glue code custom — write an intermediary server that translates between the agent and the Meta WhatsApp Cloud API.
- Proprietary plugins — wait for each vendor (ChatGPT Plugins, Claude Tool Use) to build the integration in their own way.
Result: each agent × app combination was a project. 5 agents × 20 apps = 100 integrations to maintain.
MCP breaks this by defining:
- A transport protocol (
streamable-httpeitherstdio) on JSON-RPC 2.0. - A manifest format public that describes the server and its tools.
- A standard tool definition scheme that any agent can consume.
Any MCP-compatible agent can communicate with any MCP server. A single integration.
Who supports MCP in 2026
| Client / Agent | MCP Support | Config |
|---|---|---|
| Claude Desktop (Anthropic) | Native since 2024 | claude_desktop_config.json |
| Claude Code (CLI) | Native | Claude MCP Add |
| ChatGPT / Codex CLI (OpenAI) | Native since 2025 | ~/.codex/config.toml |
| Cursor | Native | ~/.cursor/mcp.json |
| VS Code + GitHub Copilot | Native since 2025 | github.copilot.mcp.servers |
| Windsurf, Continue.dev, Zed | Native | Editor's own configuration |
| Custom agents (Python/Node SDK) | Via official SDK | SDK MCP Client |
| Gemini (Google) | Partial (via Vertex AI) | Evolving |
By the end of 2026, virtually every serious agent will natively support MCP.
Anatomy of an MCP server
An MCP server exposes three things:
- Public Manifesto in
/.well-known/mcp.json(or equivalent). Describe the name, vendor, authentication, and recommended configuration for the main clients. - Endpoint MCP (typically
/mcp) that speaks JSON-RPC 2.0. There the agent lists tools (tools/list) and invokes them (tools/call). - Tools scheme: each tool has a name (
search_chats), description, typed parameters (with a description of each one) and a return value.
When the agent connects, it reads the manifest, lists the tools, and from that moment on can invoke them as if they were local functions.
What does MCP mean for WhatsApp Business?
WhatsApp is the #1 messaging channel in Latin America and emerging markets. Connecting an AI agent to WhatsApp typically requires:
- Account with a BSP (Business Solution Provider) authorized by Meta.
- API key with permissions on the number and template.
- Incoming webhook logic for receiving messages.
- Outgoing sending logic (text, templates, multimedia).
- CRM in place to avoid losing contacts or history.
- Rate limits to avoid degrading the quality of numbers in Meta.
With an MCP server in front of all that, the agent only invokes toolsShe doesn't care how Aurora talks to Meta; she sees send_message, search_chats, create_contact and calls them. The MCP server handles auth, rate limit, audit log, multi-channel, and CRM.
Comparison: MCP vs alternatives
| Dimension | MCP | Webhooks Only | Custom API + glue |
|---|---|---|---|
| Setup time | 5 minutes | 1-2 days | 1-2 weeks |
| Lines of code | 0 | ~200 | 500+ |
| Multi-agent | ✅ Native | ❌ Each agent reimplements | ❌ Each agent reimplements |
| Audit trail | ✅ Per server | Manual | Manual |
| centralized rate limit | ✅ | ❌ | Manual |
| Vendor lock-in | Under (open protocol) | Under | High |
| Discoverability | ✅ Public Manifesto | ❌ | ❌ |
| Maintenance | The server evolves, the agent does not. | Each agent | Each integration |
Aurora Inbox MCP server
Aurora Inbox publishes its MCP server on https://developers.aurorainbox.com/mcp with public manifesto in https://developers.aurorainbox.com/.well-known/mcp.json. Supports:
- MCP Protocol version
2025-03-26. - Transport
streamable-http. - Auth via Platform API key (
Authorization: Bearer ak_live_…). - 30+ tools that cover chats, messages, contacts, companies, opportunities, channels, chatbots, teams, users, webhooks, and diagnostics.
- Multi-channelThe same tool works for WhatsApp Cloud API, WhatsApp Business, Messenger, and Instagram (
channelType0-3). - Rate limits per tenant in
send_message(1 message/sec) to protect number quality with Meta. - Native audit log: each call is associated with the
keyIdthat originated it.
This means: any MCP-compatible agent can drive Aurora end-to-end with a single key. The same key works in Claude Code, Codex, Cursor, Claude Desktop, and VS Code Copilot.
Typical use cases
Automated triage
An agent checks unanswered chats every morning (search_chats(unanswered=true)), prioritizes by urgency and leaves an internal note on each one with the response plan.
Proactive outreach
An agent connected to your CRM detects leads who opened an email but didn't reply, opens or reuses a WhatsApp chat (create_chat) and sends a personalized follow-up (send_message).
Programmatic onboarding
An agent receives an Excel file with 500 new contacts, inserts them into Aurora with upsert_contact and opens opportunities for those who qualify for BANT.
Hybrid operations
Front-line: Aurora's chatbot using GPT-5 handles the 70% volume. Back-office: Claude Code on the rep's machine automatically escalates flagged chats. unanswered=true after 5 minutes.
Common mistakes when adopting MCP
- Confusing an MCP with an LLMMCP is just the protocol; the LLM (Claude, GPT, Gemini) is still what reasons.
- Thinking that MCP replaces the REST APIAurora MCP mirrors the Platform V1 REST API 1:1 — they are two ways of talking to the same thing.
- Do not revoke keys of devs who are leaving. Each
ak_live_*It must be by human or machine, and individually revocable. - Bypass rate limits.
send_messageIt's limited to 1/sec per tenant in MCP, just like in REST. It's not a bug; it's meta-number quality protection.
What do you gain by adopting MCP in your stack?
- Your sales team uses Claude Desktop to respond to WhatsApp messages with CRM context.
- Your support team uses Cursor to resolve issues by writing prompts in natural language.
- Your marketing team uses Codex CLI to run massive campaigns respecting rate limits.
- Your product team uses VS Code Copilot to build integrations without writing code.
- One single key, Aurora, all agents.
Frequently Asked Questions
Is MCP only from Anthropic?
No. Anthropic released it as a public spec in 2024. OpenAI, Google, Microsoft, and most IDEs adopted it in 2025-2026.
Do I need to program to use MCP?
Not for already supported clients (Claude Desktop, Claude Code, Codex, Cursor, VS Code). You just paste a config snippet. For custom agents, there are SDKs in Python, Node, Go, and Rust.
Does MCP work directly with WhatsApp?
Not directly with Meta's WhatsApp Cloud API. You need an MCP server on top of that—Aurora Inbox provides one ready to use, authenticated, and with CRM support.
What happens if my MCP agent makes a mistake?
Aurora validates each call (required fields, valid GUIDs, user roles) and returns typed errors. Your agent receives the error and can retry or escalate.
Does MCP support response streaming?
Yes. Transportation streamable-http It allows the server to push events to the agent — useful for tools that take time (e.g., starting a chatbot, syncing a catalog).
How do I begin?
Create an account Aurora InboxGenerate your key in the Developer Portal and paste it into your favorite MCP client. 5 minutes from start to finish. Specific guides:
