Cold Email MCP Server: Let Claude or ChatGPT Run Warm-Up Safely

A cold email MCP server lets an AI assistant manage warm-up and outbound tasks through 28 account-scoped tools, using the same BobWork API key. If you ignore permissions and caps, an agent can repeat operator mistakes faster. This guide shows the setup, safe prompts, limits, and review checks before you let Claude, ChatGPT, Cursor, or another MCP client act.

BobWork’s MCP endpoint is for operators who manage multiple inboxes, DNS records, and campaigns. Instead of asking an assistant to reason from screenshots, you let it call approved tools. It can check the same warm-up and campaign state you see in the free warm-up tool, then summarize what needs attention.

The useful part is not novelty. It is repeatable inspection. An assistant can check every mailbox, group failures, review SPF, DKIM, and DMARC, and prepare a campaign brief before you approve sending.

What is MCP?

MCP means Model Context Protocol. It is a standard way for an AI assistant to call external tools, read structured results, and take actions you allow, instead of guessing from pasted text.

The official Model Context Protocol documentation describes the pattern: a client connects to a server, the server exposes tools, and the model asks to use those tools when a task needs live data or an action.

For email operations, that matters because stale information is risky. A mailbox can be healthy on Monday and blocked on Tuesday. DNS can pass for one domain and fail for another. A campaign can look ready until one sender crosses a bounce threshold.

An MCP server does not make the assistant better at deliverability by itself. It gives the assistant controlled access to current account data and permitted actions.

Use a cold email MCP server

BobWork’s remote MCP server is available at:

https://app-emailwarmup.bobwork.ai/mcp

It uses the same Bearer key model as the BobWork API. Personal API keys start with bwk_, and an account can have up to 5 keys.

The server exposes 39 MCP tools for AI assistants and agents, including Claude, ChatGPT, Cursor, and other MCP clients. Those tools can check warm-up, connect supported mailboxes, inspect DNS checks, and run Outbound campaigns through your account.

Use it for jobs like these:

• Find mailboxes with blocked logins or permission failures.
• List senders with weak health scores.
• Check whether SPF, DKIM, and DMARC pass for each domain.
• Confirm which mailboxes are ready for outbound.
• Create a campaign from your CSV structure and draft steps.
• Start a campaign only after caps and readiness rules pass.

This is not a separate sending engine. It is a remote control for BobWork. The same warm-up engine still runs in the background, and the same campaign rules still apply.

If you want the non-MCP workflow first, read the warm-up and sending from the same mailbox guide. The MCP server is best after you understand the normal operating model.

What limits matter?

Put the limits in front of the assistant. It should reason with caps, not vague sending advice. These are the numbers that affect MCP-driven warm-up and outbound work in BobWork.

The warm-up network is shared by default, or you can use a private pool per account if you prefer. In shared mode, one mailbox is enough to start because it warms with other active mailboxes in the network. In private mode, switch in Mailbox pool → Pool mode; at least two mailboxes are needed.

The free pool does the warm-up work automatically: it exchanges short plain-text office notes, replies to a share of messages, rescues its own warm-up mail from spam, and never reads, moves, or answers real mail. You can see the engine behaviour in the strategy page.

How do you set it up?

Keep setup plain. Do not use browser automation or pasted mailbox credentials. Use the remote MCP endpoint and a Bearer header.

1. Create a personal key

Create a BobWork personal API key. It will start with bwk_. Keep it out of shared documents, screenshots, CRM notes, and prompt templates.

Use one key for the MCP client you are configuring. If you later connect another assistant, create another key if you have room. The account limit is 5 keys.

2. Add the remote server

In your MCP client, add a remote MCP server. Use this server URL:

https://app-emailwarmup.bobwork.ai/mcp

Use the name your MCP client expects for remote servers. The important part is the URL.

3. Add the auth header

Add an HTTP authorization header:

Authorization: Bearer bwk_your_key_here

The exact field name varies by MCP client. The value should be sent as a Bearer token with each request.

4. Test with read-only tasks

Start with account status, mailbox health, and DNS checks. Do not start with campaign creation.

A good first test is:

List my connected mailboxes, their warm-up health score, 7-day spam rate, ready-for-outbound status, and any blocked login or permission issue.

You should see structured results from the account. If the assistant cannot see mailboxes, check the endpoint, Bearer header, and key.

5. Approve actions explicitly

Ask the assistant to prepare actions before taking them. Let it draft a campaign plan, choose eligible mailboxes, and show expected daily send volume. Then approve creation or starting.

The key can act on your account. Treat the assistant like an operator with tools, not a passive notepad.

Which prompts work best?

Good prompts are specific, limited, and verifiable. They tell the assistant what to inspect, which threshold to use, and what output format you want.

Use prompts like these for warm-up operations:

Check every mailbox. Return a table with email address, provider, health status, 7-day spam rate, ready-for-outbound status, and the single highest-priority fix.

Find all mailboxes that are Blocked, At risk, or Watch. Group them by login failure, DNS failure, spam rate, or not enough warm-up history.

Show senders with 7-day spam rate above 3%. Recommend whether to pause, keep warming, or investigate DNS. Do not start campaigns.

For each domain, check SPF, DKIM, and DMARC. Tell me which record is missing or failing, and which mailbox depends on that domain.

Find mailboxes that have warmed for at least 14 days, received at least 30 warm-up emails in 7 days, and have spam rate at or below 10%. Return only eligible outbound senders.

These prompts work because they match real product states. BobWork health score is 0–100. Healthy is 85 or above, Watch is 60–84, At risk is below 60, and Blocked means login or permission failed.

For DNS, do not let the assistant treat SPF, DKIM, and DMARC as vague advice. Ask it to check record status per domain. BobWork checks SPF, DKIM, and DMARC daily and shows results per domain.

For a deeper DNS repair workflow, keep the assistant aligned with SPF, DKIM, and DMARC for cold email. Your records must match the provider you send through.

Google’s sender guidance says spam rates should stay below 0.10% and avoid ever reaching 0.30%. Yahoo’s sender guidance requires SPF and DKIM, DMARC at least p=none, alignment, and spam rate under 0.3%. Use the primary pages from Google and Yahoo when you set policy for larger programs.

How should campaigns run?

BobWork Outbound is beta. It sends cold email sequences from your own warmed mailboxes, not from a separate vendor mailbox. The MCP server can help create and start campaigns, but it cannot bypass the rules.

A campaign can use a CSV with email, first_name, company, and custom columns. You can write up to 4 steps with delays and merge fields such as {{first_name}} or {{company|there}}.

Only mailboxes that pass the ready gate can send. That gate requires 14 or more days warming, 7-day spam rate at or below 10%, and at least 30 warm-up emails received.

Each email includes one-click unsubscribe through a List-Unsubscribe header and link. Unsubscribes and bounces are suppressed for the account forever. A reply stops the sequence for that contact. The one-click unsubscribe standard is defined in RFC 8058.

Use tighter prompts for campaign work:

Using only ready mailboxes, draft a campaign plan for this CSV schema: email, first_name, company, role. Use 3 steps. Keep tone plain. Show the senders selected and maximum daily send volume before creating anything.

Create a campaign with 2 steps from my uploaded contacts. Use only mailboxes that are ready for outbound. Set delays between steps. Do not start until I approve the summary.

Before starting this campaign, check bounce history, unsubscribed contacts, mailbox readiness, and daily cap. If any mailbox is in Watch, exclude it and explain why.

Start the approved campaign. After starting, report campaign status, selected mailboxes, daily cap per mailbox, and any suppressed contacts.

If you want to compare this with the normal interface, read pay-as-you-go cold email software. MCP is not a different pricing model. It is a different control surface.

Keep the agent boxed in

The main safety rule is simple: an MCP key can act on the account it belongs to. Do not give it to a client, contractor, shared assistant, or workspace unless you are comfortable with that access.

Use a separate key for each MCP client when possible. If something looks wrong, revoke that key and create a new one. Do not paste keys into prompts. Put them in the MCP client’s secure server configuration.

Campaigns still obey BobWork’s caps. In the first week after a mailbox becomes ready, it can send up to 30 cold emails per day. After that, it can send up to 50 per day. Sends happen only in the mailbox’s working hours, with 4–9 minutes between sends from one mailbox.

The agent also cannot send from unready mailboxes. Warm-up keeps running on the same mailboxes while they send, but the ready gate still controls access to Outbound.

Watch for three failure modes.

First, the assistant may over-select mailboxes. Tell it to exclude Watch, At risk, and Blocked senders unless you explicitly approve them for investigation only.

Second, the assistant may over-focus on warm-up and ignore consent, list quality, and relevance. Warm-up does not make a bad list safe. Keep bounces below the stop threshold and avoid purchased or scraped lists that do not match your sending policy.

Third, the assistant may create copy that sounds plausible but violates your offer or compliance rules. Treat AI copy as a draft. Review it before starting.

If replies matter more than sending volume, combine MCP campaign checks with a human reply workflow. The unified inbox for cold email explains how to manage replies across connected mailboxes without letting warm-up mail pollute the inbox.

Use this checklist before you let any assistant run operational email tasks. The first session should be boring.

If you are setting this up for the first time, start with warm-up health and DNS only. After one clean session, let the assistant prepare a campaign draft. After that, let it start approved campaigns.

For sending volume policy, pair the MCP workflow with how many cold emails per day is safe per mailbox. For new senders, add patience. BobWork recommends 14 days for an existing domain and 3–4 weeks for a new one.

BobWork Email Warm-up is free for up to 10 mailboxes, with a shared warm-up network by default and a private pool per account if you prefer. Use the MCP server when you want an assistant to check health, verify DNS, and operate approved Outbound campaigns without skipping the safety gates.

Frequently asked questions