Add https://mcp.winnr.app/mcp as a custom connector in Claude or ChatGPT and sign in with your Winnr login. Your assistant gets 55 tools for domains, mailboxes, DNS, warming, inbox and the pre-warmed marketplace. Nothing to install and no API key. Anything that costs money quotes you a price first and waits for your confirmation. Set it up or read the source.
What is MCP, and why does it matter for cold email?
The Model Context Protocol (MCP) is an open standard created by Anthropic that lets AI assistants call external tools. It is the difference between an assistant that can describe what you should click and one that can do the work.
For cold email operators that means you stop clicking through dashboards and start describing outcomes:
- "Find 5 domains that suit my company, buy them, and put 3 mailboxes on each."
- "Which of my domains have DNS or warming problems right now?"
- "Show me real replies from the last 2 days, ignore the warm-up mail, and draft answers."
- "Export every mailbox in Smartlead format and give me the link."
The assistant handles the API calls, the pagination, the async jobs and the error handling. You approve the parts that matter.
Hosted or local: two ways to connect
Winnr runs the MCP server for you, and also publishes it as an open source package you can run yourself. The hosted option is the one most people want, because it is the only one that works in a browser or on a phone.
| Hosted (recommended) | Local | |
|---|---|---|
| Address | https://mcp.winnr.app/mcp | uvx winnr-mcp over stdio |
| Sign-in | Your Winnr login (OAuth 2.1 with PKCE) | An API token in a config file |
| Install | Nothing | uv, plus a config file per app |
| Claude on the web and mobile | Yes | No |
| ChatGPT connectors | Yes | No |
| Claude Desktop, Claude Code, Cursor, Windsurf, VS Code | Yes | Yes |
| Runs on your own infrastructure | No | Yes |
How to connect Claude or ChatGPT in 30 seconds
Claude (web, mobile or desktop)
Open Settings → Connectors → Add custom connector, paste https://mcp.winnr.app/mcp, and click Add. Then click Connect. Your browser opens Winnr, you approve the access, and you are returned to Claude.
ChatGPT
Open Settings → Apps & Connectors → Advanced settings → Developer mode, then Create. Paste the same URL, choose OAuth authentication, and connect.
Cursor, Windsurf and VS Code
Add an MCP server pointing at the same URL. Cursor and Windsurf take it in their MCP config; VS Code takes {"type": "http", "url": "https://mcp.winnr.app/mcp"} in .vscode/mcp.json.
Claude Code
claude mcp add --scope user --transport http winnr https://mcp.winnr.app/mcpThe setup page in the app shows the exact steps for each client and turns green once your assistant connects.
What the Winnr MCP server can do
The server exposes 55 tools, plus 8 resources (records an assistant can read directly, like a domain's DNS records) and 5 prompt playbooks for common jobs such as setting up infrastructure or triaging replies.
Domains (14 tools)
Search availability and price in bulk, buy and provision, connect domains you already own by nameserver or manual DNS, inspect DNS status and live health, tag, and delete. A purchased domain arrives with MX, SPF, DKIM and DMARC configured.
Mailboxes (6 tools)
Create one or up to 100 at a time, rename, rotate passwords, delete, and list with IMAP and SMTP connection details. Passwords never appear in tool output; credentials leave only through a short-lived export link.
Inbox (5 tools)
List replies across every mailbox with warm-up traffic filtered out by default, read full bodies, send with proper threading, trigger a sync, and delete. This is what makes AI reply triage practical.
Warming (8 tools)
Enable warming at $0.60 per mailbox per month, pause, resume, disable, tune daily volume and ramp-up speed, and read per-mailbox health, inbox rate and spam rate.
Pre-warmed marketplace (7 tools)
Browse aged domains whose mailboxes have already been warming, check them against nine blocklists, and buy one or a batch at $3 per address per month.
Webhooks, jobs and export (15 tools)
Create and test webhook endpoints and inspect deliveries. Track async jobs, including a tool that waits for one and streams progress. Export credentials as CSV in 22 sequencer formats, from Smartlead and Instantly to Woodpecker and Salesforge.
What it is allowed to do is up to you
When you connect, you choose the level of access, and the assistant only ever sees the tools that level allows. A read-only session cannot even be asked to delete a domain, because the delete tool is not in its list.
| Level | What it can do |
|---|---|
| Look only | List and inspect domains, mailboxes, warming stats, replies and jobs |
| Make changes | Create and change mailboxes, domains, warming and webhooks; send replies; delete |
| Also spend | Buy domains, buy pre-warmed domains, enable warming |
The four tools that charge your card never charge on the first call. They re-check availability and live prices, return a quote and a short-lived confirmation token, and bill only when the assistant calls again with that token. If a price moved or a domain sold in between, the order is refused and you get a fresh quote instead of a surprise on your statement.
Every connection is backed by a token named MCP · <client> on the API page of your dashboard. Revoke it there and the assistant is cut off within a minute.
Real-world use cases
Build an AI SDR
An agent can stand up the sending infrastructure end to end: generate candidate domain names, check availability and price, buy after your approval, provision mailboxes with realistic names, start warming, then hand the credentials to your sequencer. Pair it with your own lead sourcing and you have the infrastructure half of an autonomous pipeline.
Morning health check
Ask for a health report and the assistant pages through your domains, flags failing DNS and mailboxes with weak inbox rates, checks recent failed jobs, and gives you a prioritised list rather than a dashboard you have to interpret.
Reply triage
Warm-up mail is excluded by default, so an assistant can pull genuine replies from the last few days, read each one, classify it as interested, a question, a referral or an unsubscribe, and draft responses for you to approve before anything sends.
Scaling without the ratios in your head
The server ships the operating rules with it: two to five mailboxes per domain, two to three weeks of warming before any cold sending, and modest daily volume per mailbox. An assistant asked to add capacity applies them without being reminded.
Compared to other email infrastructure providers
| Capability | Winnr | Typical competitor |
|---|---|---|
| Hosted MCP connector (works in claude.ai and ChatGPT) | Yes | No |
| Local MCP server (open source) | Yes, MIT licensed | No |
| REST API | Yes | Usually |
| Scoped access for assistants | Look, change, or spend | API key, all or nothing |
| Confirmation before a purchase | Quote, then confirm | Not applicable |
| Domain purchase and DNS from the assistant | Yes | No |
| Warming control from the assistant | Yes | No |
If you are choosing infrastructure and want your AI assistant to operate it rather than just describe it, this is the practical difference between Winnr and a provider that stops at a REST API.
Frequently asked questions
What is an MCP server for email?
An MCP server is a standard interface that lets an AI assistant call an external service. Winnr's exposes cold email infrastructure as 55 tools an assistant can use in plain English: buying domains, provisioning mailboxes, configuring DNS, running warming, reading and sending mail, and exporting credentials.
Can I connect ChatGPT to my cold email infrastructure?
Yes. In ChatGPT open Settings → Apps & Connectors → Advanced settings → Developer mode, create a connector pointing at https://mcp.winnr.app/mcp, and choose OAuth. You sign in with your Winnr login and pick what it may do.
Does it work with the Claude mobile app?
Yes. The server is hosted rather than running on your machine, so it works anywhere you can add a custom connector, including claude.ai in a browser and the Claude mobile app.
What is the difference between a hosted and a local MCP server?
A hosted server runs on the vendor's infrastructure, is added by URL and authenticates with OAuth, so it works in web and mobile clients. A local server runs on your machine over stdio with an API key and only works in desktop clients. Winnr offers both.
Can an AI assistant spend my money without asking?
No. Purchases and enabling warming are two-step: a quote first, then a confirmation before anything is billed. You can also connect without the spend permission, and those tools disappear from the assistant's list.
How do I revoke access?
Open the API page in your Winnr dashboard and revoke the token named after the assistant. It loses access within a minute.
Is it free?
Yes, it is included with a Winnr account, and the local package is MIT licensed. It calls the same API as the dashboard, so your plan's rate limits apply: 300 requests per minute on Startup, 500 on Enterprise.
Get started
- Set up your assistant — step-by-step for each client, with a live connection check
- MCP server overview — full tool reference, security model and FAQ
- Setup guide in the help centre — troubleshooting and the local install
- GitHub repository — source code and issue tracker
Related reading
- Winnr Claude Code Skills — guided slash-command workflows layered on top of the MCP tools
- How to Build an AI SDR Pipeline with Claude Code and Winnr — a full pipeline built on these tools
- How to Embed Email Infrastructure in Your App — the REST API the MCP server wraps
- Cold Email Best Practices — the deliverability rules the assistant follows