---
title: "Sumora MCP Server: Run Outbound From Claude, ChatGPT or Gemini"
url: https://trysumora.com/connect/sumora-mcp-server
markdown-of: https://trysumora.com/connect/sumora-mcp-server
---

# Sumora MCP Server: Run Outbound From Claude, ChatGPT or Gemini

[Home](/)/[Connect](/connect)

Connect

# The Sumora MCP server

Add Sumora to Claude, ChatGPT, Gemini, Cursor or Windsurf. Review approvals, read the unified inbox, query your business brain, and delegate to agents — with send ceilings still enforced in the engine.

By Devansh Ranjan · Published Aug 7, 2026

**Server URL:** `https://mcp.trysumora.com/workspace/mcp` — add it as a connector in Claude, ChatGPT, Gemini, Cursor, or Windsurf and authenticate with a workspace key. Send ceilings stay enforced in the engine and cannot be raised from the client.

Most outreach MCP servers are a remote control: the assistant asks, the server does. That is fine until the assistant asks for something the account cannot survive. Sumora’s splits the two concerns — the agent runs outbound continuously on its own, and MCP is how you supervise it.

## What the server exposes

-   **Pending approvals** — review and approve queued outreach without leaving the chat.
-   **Unified inbox** — LinkedIn, Instagram, and email replies in one place.
-   **Business brain** — query what Sumora knows about your company, ICP, and pipeline.
-   **Agent delegation** — hand work to specialist agents.
-   **Notifications** — what needs attention now.

## Setup

Full walkthrough at [trysumora.com/mcp](/mcp). In short: copy your workspace key from Settings → API Keys, add `https://mcp.trysumora.com/workspace/mcp` as a custom connector, and authenticate when prompted.

## What it deliberately will not do

The ceilings in the [capability table](/capabilities) are enforced in the engine. That means:

-   An assistant cannot raise a per-account daily or weekly send limit.
-   Cold first touches still require approval; there is no autonomous-send flag.
-   Instagram cold DMs stay blocked until the prospect has been warmed — 2 touches over 18+ hours.
-   Sending stays inside the 08:00–20:00 weekday window.

This is a deliberate design choice rather than a missing feature. A connector that obeys any volume request transfers the account risk to you — see [what actually triggers a restriction](/guides/linkedin-connection-request-limits) and [why approval is a rate control](/guides/ai-outbound-approval-architecture).

## Step by step

1.  **Get your workspace key.** In the Sumora dashboard, open Settings then API Keys and copy your workspace key.
2.  **Add the connector.** In Claude, ChatGPT, or Gemini connector settings, add a custom MCP server with the URL https://mcp.trysumora.com/mcp
3.  **Authenticate.** Paste the workspace key when the client prompts you to log in.
4.  **Confirm the tools loaded.** Ask the assistant to list your pending approvals. A response means the connection is live.
5.  **Operate from chat.** Review and approve queued messages, read replies, query the business brain, and delegate work to specialist agents.

## Frequently asked questions

What is the Sumora MCP server URL?

https://mcp.trysumora.com/mcp — add it as a custom connector in your AI client and authenticate with a workspace key from Settings then API Keys in the Sumora dashboard.

Which AI clients work with the Sumora MCP server?

Claude, ChatGPT, and Gemini through their connector settings, plus Cursor, Windsurf, and other MCP-compatible clients. MCP is a shared protocol, so one server URL works across all of them without reconfiguration.

Can I send LinkedIn or Instagram messages directly from Claude through MCP?

You can review and approve messages the agent has already drafted and queued. You cannot bypass the per-account send ceilings or the approval requirement, because both are enforced in the Sumora engine rather than in the client. An assistant asked for 200 connection requests still gets the seat's daily cap.

Why enforce limits in the engine instead of the client?

A chat session has no memory of how many invitations the seat already sent this week, so a client-side limit is not a limit at all. Enforcing in the engine means the ceiling holds no matter which assistant, session, or user is asking.

----

_Markdown twin auto-generated from the HTML page at build time. Source of truth: the HTML/_next data._