Skip to main content

​
What is MCP?

Model Context Protocol (MCP) is an open standard that lets AI assistants connect securely to business tools. Instead of copying data into a chat, you connect CloudTalk once. From then on, the assistant sees a set of CloudTalk tools (search calls, read a transcript, update a contact) and picks the right one for whatever you ask. Every request is authenticated, by your CloudTalk sign-in or your API key, and scoped to your account, and CloudTalk returns only the data that one tool call asked for.

​
What you can do

Build AI Voice Agents

Create and tune voice agents, manage their knowledge bases, and review how their calls went.

Analyze conversations

Search call history, pull transcripts, read AI summaries, sentiment, topics and smart notes.

Understand customers

A 360° view of any contact: every call, note and activity in one timeline.

Act on your account

Create and update contacts, tag calls and contacts, send SMS, place calls. Your AI app asks you to approve them before they run.

​
Profiles

Assistants pick tools more reliably from a short, focused list. Connect the profile that matches the job instead of the full catalogue:

​
Common use cases

See the Recipes page for ready-to-use prompt chains, grouped by job: coaching and QA, live-call assist, AI Voice Agents, and contact data.

​
Your AI app asks before actions run

Every tool acts on its first call: there is no separate preview step. Before a tool that sends, dials, or changes data runs, your AI app shows an approval prompt with what it is about to do. The tools marked Asks every time in the tool reference reach a person (an SMS, a call, a campaign start) or overwrite or delete data, so Claude and ChatGPT ask before each call by default. An admin or a user can change that and let a tool run without asking: keep approval on for these tools. The approval is your app’s, not CloudTalk’s. A client without approval prompts, such as an automation tool or your own code, runs the action as soon as it calls the tool. Choose an assistant and a client whose approval prompts you trust. Nothing in the catalogue can delete contacts, calls or recordings.

​
Security

  • Every tool is scoped to your company by the account you signed in with or by the API key. No tool accepts an account ID, so an assistant can’t reach another company’s data.
  • Signing in with CloudTalk is open to users with the Admin role only, and you approve the access on a consent screen before the assistant gets it.
  • CloudTalk sees only tool parameters (a date range, a call ID), never your conversation with the assistant.
  • Recording links are signed and expire within 12 hours, often sooner, and every access is logged in CloudTalk’s audit trail.
  • To cut off access, disconnect the connector in your assistant if you signed in, or revoke the API key in the Dashboard; key revocation can take a few minutes to take effect.

​
Availability

MCP is available on the Essential plan and higher. Two areas depend on what else is on your plan:
  • AI insights (summaries, sentiment, topics and transcripts) require Conversation Intelligence. Calls without it return empty AI fields rather than an error.
  • Campaign tools require the new Sales Dialer add-on.

​
Limits and freshness

  • 60 requests per minute per company. A 429 means wait a minute and retry.
  • Reads are cached briefly (seconds to minutes) to keep responses fast; live queue stats are never cached. Records you change through the assistant update immediately, though a newly created voice agent or knowledge base can take a short while to appear in list results.
  • AI insights need a recorded, processed call. A call that hasn’t been analysed yet returns empty AI fields, which is a normal answer.

​
FAQ

MCP is available on the Essential plan and higher. AI insights (summaries, sentiment, topics, transcripts) additionally require Conversation Intelligence, and the campaign tools require the new Sales Dialer add-on. If you’re on Starter, upgrade your plan to get access.
There are two ways to connect. Sign in with CloudTalk (OAuth) works in Claude (Claude.ai, Claude Desktop, Claude mobile), Claude Code, ChatGPT and Gemini. You sign in as a CloudTalk admin and approve the access, with no key to copy. An API key in a request header works in every client that can send one: Claude Code, Cursor, and any client that supports Streamable HTTP with a custom header. In Claude.ai it needs the request-header field, a beta Anthropic enables per organization. The Quickstart has the setup for each.
It can’t delete a contact, call or recording. It can still remove data through a write: updating a contact or replacing a campaign’s sources replaces the tags, numbers or emails you don’t include, removing call tags detaches them, and deleting a campaign is terminal: archived is the reversible option, deleted cannot be undone. By default, your AI app asks you to approve the actions that carry the most weight every time; they’re marked Asks every time in the tool reference.
Only the tool calls: a date range, a call ID. Your conversation stays with your AI provider; whether the assistant stores or trains on your chats is governed by that provider’s settings, not CloudTalk’s.
Signed in with CloudTalk, the assistant reaches the company of the account you signed in with, and each member of a Claude organization signs in separately. With API keys, one key is scoped to one company, so two accounts means two keys: in clients you configure yourself (Claude Code, Cursor) register one server per account. A Claude connector holding a header acts for everyone as the single account whose key it holds, and Claude refuses a second connector on a URL that already exists in your organization.

​
Next steps

Quickstart

Connect Claude, Cursor, or another MCP client in five minutes.

Tool reference

Every tool with its parameters, defaults and limits.