Stop replying to leads by hand. Run Instantly from Claude.

What you get once this is connected

MCP (Model Context Protocol) is the cable between Claude and Instantly. Once it is plugged in, Claude can read your account live and act inside it from a normal chat. No CSV exports, no tab switching.

• Pull every reply across campaigns and separate real interest from auto-responders and rejections.
• Read a full thread, both what they sent and what you sent, before a single word gets drafted.
• Judge fit against your own rules and tell you to skip a lead instead of wasting a follow-up on it.
• Draft the reply that continues the conversation instead of repeating a pitch you already sent.
• Audit campaign health: reply rates, bounce rates, sending accounts, warmup, DNS records.
• Build and edit sequences, manage leads and lists, and run blocklist hygiene.

Before you start

• An Instantly account with API access on your plan.
• A Claude account. Custom connectors are available on paid plans.
• Admin rights on the Instantly workspace you want to connect.
• Ten minutes and somewhere safe to store a key.

Step 1 · Generate your Instantly API key

1. Log in to Instantly.
2. Open Settings, then the Integrations or API section.
3. Choose API keys and click Create key (use API v2).
4. Name it something you will recognise later, for example claude-mcp.
5. Set the scopes. Start read-only, and add write scopes later once you trust the setup.
6. Copy the key immediately. Most platforms show it once.

While you are on that page, look for the MCP server URL Instantly provides. You will need it in the next step.

Step 2 · Connect the MCP to Claude

Two methods. Method A is the one most people should use.

Method A · Custom connector in Claude settings

1. In Claude, open Settings and find Connectors.
2. Click Add custom connector.
3. Name it Instantly MCP.
4. Paste the MCP server URL from Instantly and replace the placeholder with your own API key.
5. Click Add. The connector goes live, and you can open Configure to control which actions Claude is allowed to perform.

https://mcp.instantly.ai/mcp/<YOUR_CREDENTIAL_STRING>

# Instantly gives you this URL on the API page.
# The credential portion IS your secret. Treat the whole URL as a password.

Method B · JSON config (Claude Desktop and Claude Code)

If you prefer to declare the server in a config file, use the header form so the key is not sitting in a URL path.

{
  "mcpServers": {
    "instantly": {
      "url": "https://mcp.instantly.ai/mcp",
      "headers": {
        "Authorization": "PASTE_YOUR_API_KEY_HERE"
      }
    }
  }
}

Save, then fully restart the Claude app so the server is picked up.

Turn off what you do not need

Open the connector's settings and disable the destructive tools until you actually want them: anything named delete, plus campaign activation and send actions. You can enable them per task later.

Step 3 · Verify the connection

Do not start with a real task. Start with something harmless that proves the pipe works.

If you get a list back with real campaign names, you are connected. If you get an authentication error, jump to troubleshooting.

Step 4 · The tool map

The connector exposes a large tool set. You do not need to memorise it. These are the ones that carry almost all real work.

The important parameter to learn is on list_emails. It takes an email type, and that single argument is where most people's workflow quietly breaks.

Step 5 · The blind spot that breaks everything

Here is the part nobody tells you, and it cost me a nearly embarrassing email before I caught it.

When you pull messages with the type set to received, you get only what the lead sent you. Your own replies are a separate dataset. So if you ask Claude to "read the thread" and it only pulls inbound mail, it sees their questions and never sees that you already answered them three days ago.

The result is a follow-up that re-answers a solved question. To the lead, it reads like you do not read your own emails.

What Claude sees with inbound only, versus what actually happened:

The fix

• Pull both types, sent and received, for the campaign.
• Stitch them together on the thread ID, which is the reliable key. Sort by timestamp.
• Identify who sent the true last message. That single fact decides the whole reply:
  • Last message was theirs → answer what they asked.
  • Last message was yours → nudge your own message, do not re-pitch.

Step 6 · Set the safety rules first

Give these rules once at the start of a working session. They are the difference between a useful assistant and a liability with access to your sending accounts.

• Never let it auto-send. One badly judged email under your name is worse than a hundred unsent ones.
• Keep destructive tools off unless you are actively using them.
• Re-state the rules if the conversation runs long. Long sessions drift.

Step 7 · The reply workflow, lead by lead

This is the loop that actually produces booked meetings. Work one lead at a time. Batching feels faster and produces generic mail.

1. Pull the replies. Every inbound message across active campaigns.
2. Sort by real interest. The reply data carries an interest signal, positive, neutral, negative. Auto-responders and rejections drop out here.
3. Remove anyone already booked. The lead status field carries this. Do not follow up on a meeting you already have. This one step cut my list nearly in half.
4. Read the full thread for the lead in front of you, both directions.
5. Judge fit before drafting. See the next section. A skip is a result.
6. Draft, review, send manually, then move to the next lead.

Step 8 · Teach it your fit rules

An assistant that drafts a beautiful email to a lead you should have dropped is not helping you. Give it your disqualifiers explicitly, in writing, and let it call the skip.

These are anonymised patterns from real threads in my own campaigns. Yours will differ, but the shape holds.

Step 9 · Prompt library

Copy these. They are the ones I actually reuse.

Step 10 · Troubleshooting