MCP or a webhook: which one you actually need
MCP is request and response: an assistant calls a tool when a person asks it to. A webhook is the opposite shape, where Meta delivers an event and your system decides what to do. Nothing that reacts automatically can be built on MCP alone.
- MCP acts, webhooks react
- No assistant wakes at 3am
- MCP can configure the webhook
Nothing charged for 7 days. Cancel anytime.
Integration paths a connected number offers: MCP and webhook.
Ways an incoming message can trigger an MCP server.
The MCP tool that configures the other path.
Why can MCP not answer a customer automatically?
Because the protocol has no mechanism for being woken.
An MCP server exposes tools that a client calls. The call originates with the assistant, which is itself responding to a person. There is no inbound direction: nothing about a customer sending a WhatsApp message can reach into an assistant and cause it to act. This is true of every MCP server in every product, not a limitation of ours.
So the question every buyer eventually asks, whether an AI can reply to customers overnight, has a webhook answer rather than an MCP one. Meta delivers the message to a system you run, and that system decides what to do, which may well include calling a model.
What is each path genuinely good for?
MCP is for the work a person is doing. Which of my numbers is unhealthy, send this customer their confirmation, create a template for order updates, point this client's number at my server. Each of those starts with somebody asking, and each is faster through an assistant than through a dashboard.
The webhook is for everything that has to happen without anyone present. First reply after hours, routing to the right person, logging to a CRM, triggering a workflow.
The useful property is that MCP can set up the webhook. get_connect_link produces the link a client opens, and set_webhook_destination points their number at your system, so an agent can do the onboarding for the path it cannot itself serve.
Which tools speak which?
Assistants speak MCP: Claude, ChatGPT, Cursor, Claude Code. Automation platforms mostly do not, and the distinction is not what people expect.
Make and Zapier both run MCP servers of their own, exposing their actions to assistants. Neither acts as an MCP client, so neither can call out to ours. For them the integration is a webhook and an HTTP call.
n8n is the exception. Its MCP Client Tool connects out to a remote server and supports OAuth2, so an agent node inside an n8n workflow can call our tools directly. Even there the workflow still starts with a webhook, because that is the only thing that reacts.
Common mistakes
- Buying MCP expecting automatic replies. It acts when asked and nothing asks it at 3am.
- Assuming a platform with an MCP server can consume one. Make and Zapier expose, they do not consume.
- Building both paths to do the same job. They are for different halves of the same system.
The connector includes set_webhook_destination, so the assistant path can configure the automation path without anyone opening a dashboard.
Frequently asked questions
Can an assistant watch my WhatsApp?
No. Nothing can push an event into an MCP server. Watching is a webhook, and the webhook can then call a model.
Do I need both?
Most real systems use both: MCP for the work people do, the webhook for everything that happens without them.
Can Zapier use your MCP server?
No. Zapier is an MCP server rather than a client, so it has no way to call out to ours.
Can n8n?
Yes, through its MCP Client Tool with OAuth2. It is the one automation platform of the three that can.
Keep reading
Ready to get started?
Set up WhatsApp Coexistence in minutes, not months. The app keeps working on the phone.
Start Free TrialNothing charged for 7 days. Cancel anytime.Verified on