Skip to main content
The MCP server gives an AI app what the API gives your code. It finds people and companies, saves them to lists, finds work emails and phone numbers, and runs email and LinkedIn campaigns. Every tool is one API call: it is checked, charged and logged exactly as that call is.
The server speaks Streamable HTTP. An app signs in with OAuth, and a member of your organization approves what it may do. A client that cannot sign in can send an API key instead.

Connect an app

Owners, admins and developers can connect apps.
1

Add a custom connector

In Claude, open Customize > Connectors, select Add, then Add custom connector.
2

Name it and paste the URL

Name it outbound.dev, then copy and paste the MCP server URL above. Keep the sign-in settings Claude suggests, then select Add.
3

Approve it

On the outbound.dev page that opens, pick the organization and the access it gets, then select Approve.
The console shows the same steps, with a button that copies the URL, under MCP. Once an app is connected, see Work with an AI assistant for requests that work and what each one costs.

What an app may do

The member who approves an app decides three things on the approval page: An app lists only the tools its access covers, so a read-only app is never offered a tool that changes something. Connected apps are listed in the console under MCP, each with its access, what it spent today and when it was last used. Change an app’s access or cap there, or disconnect it. A change applies from the app’s next call, and the app does not have to connect again. Owners, admins and the member who approved an app can change it. An organization can have up to 100 connected apps.

Tools

The server offers more than 100 tools. Each one is an API call. Four calls are not offered as tools yet: replying in a conversation, sending a new message, sending a LinkedIn invitation, and sending a step as a test email. Use the API for those. A tool with a long answer gives a short one by default. The app asks for everything with detail set to full.

What a call costs

A tool costs what its API call costs: see Pricing and billing. A tool that charges states its price in its description, and every answer says what that call charged. Reading is free, and a lookup that finds nothing costs nothing. Every tool call is recorded like any other call, with its own request id and cost. Find it in the console under Logs, or ask the app for your recent calls.

Working safely with an assistant

The server tells every app how to use it:
  • Ask first. An app confirms with you before a tool that spends, sends or deletes. Each tool is marked as read-only or not, so an app that asks for approval knows when to ask.
  • Nothing sends twice by accident. A tool call that is sent again runs again, so an app is told never to repeat a send that may already have gone out.
  • Outside text is data. Text inside profiles, list rows, messages and web pages was written by other people. An app is told to treat it as data, never as instructions.
  • Ids come from answers. An app uses only ids an earlier answer gave it.
A cap is the firm limit. Set a daily cap for any app that may spend, as you would for a key.

Limits

  • A tool call counts against the same rate limits as its API call: see Rate limits. Listing the tools counts against the organization’s limit.
  • A long run, such as enriching a list, answers with a job. The app reads the job until it ends.
  • An app that is not used for 30 days signs in again.

Use an API key instead

A client that cannot sign in can send an API key as a Bearer token:
The key’s own scope and daily cap apply, as they do on the API.

Building a client

A call without a token answers 401 with a WWW-Authenticate header that names where to sign in.