Zapier (via MCP)

ThunderPhone connects to Zapier through a remote Model Context Protocol (MCP) endpoint. The endpoint advertises a set of tools, ThunderPhone discovers those tools, and an agent can call the attached tools during a phone conversation. This is useful when a call needs to trigger an action in a system that does not have a native ThunderPhone app connection.

The important boundary is that Zapier is not a fixed set of built-in ThunderPhone actions. What the agent can do depends entirely on the tools exposed by your Zapier MCP endpoint. If the endpoint does not advertise an action, that action is not available to the agent.

What the Zapier connection does

You add the endpoint under Connections → MCP, where ThunderPhone saves the remote server and synchronizes its tool list. The connection uses Streamable HTTP; a local server that works only over standard input and output cannot be attached. Once the tool list has been discovered, you attach the MCP server to a specific agent in the builder.

Attached tools are callable during that agent's conversations. The agent chooses when to call them from the conversation and its prompt. This is an example of function calling: the caller speaks naturally, the agent maps the request into a structured tool call, and the remote endpoint returns a result the agent can use in its reply.

ThunderPhone caches the discovered tool list. If you add, rename, or remove actions at the endpoint, synchronize the server again before relying on the changed set in calls.

Typical call flows

Lead handoff. In a lead qualification call, the agent asks the approved qualifying questions. If your endpoint exposes an action for the destination system, the agent can submit the collected fields before ending the call. The prompt should say which fields are required and what to do when the tool returns an error.

Structured phone intake. For a phone intake form, a discovered action can receive the caller's confirmed answers. Have the agent repeat sensitive or easy-to-mishear values before it invokes the tool, and keep a human route for requests the destination workflow does not accept.

Internal follow-up. If the endpoint exposes an appropriate task or notification action, the agent can invoke it near the end of the conversation after the caller's request is clear. It should describe the result only after the tool returns successfully; submitting a tool call is not the same as confirming the downstream work completed.

Setup outline

1. Configure the actions you want to expose through your Zapier MCP endpoint. Use narrow, clearly named actions with only the fields the call flow needs.
2. In ThunderPhone, open Connections → MCP and add a server. Give it a recognizable name and paste the remote endpoint URL.
3. Save and synchronize the server. Review the discovered tool names; an empty list means the endpoint advertised no tools.
4. Open the agent in the builder and attach the server in its MCP section. A saved server has no effect on calls until it is attached.
5. Update the agent prompt with rules for when to use each action, which caller details require confirmation, and when to transfer or stop instead.
6. Test successful results, validation failures, unavailable endpoints, and ambiguous caller answers before deploying the agent configuration.

Where MCP ends and other integrations begin

Use the Zapier MCP connection for tools an agent should deliberately call while speaking with someone. Use a webhook when your own server needs a call event notification, such as a completed-call event, regardless of what the agent decides during the conversation. That distinction keeps conversational actions separate from event delivery.

For the underlying protocol, see Model Context Protocol. Keep permissions and downstream data handling aligned with the exact information your call flow collects.