How to run outbound from Claude

Claude is already good at the hard parts of outbound: researching an account, deciding whether it fits, and writing a first line that does not read like a template. What it does not have out of the box is the rest of the job: memory across sessions, a cadence that keeps running when you are not there, and someone watching for replies.

This guide is the setup that fills those gaps. It works in Claude Code, Claude Desktop and Cowork, because each of them can connect to tools over MCP.

1. Write your qualification questions first

Before any tools, decide what makes an account worth contacting, written as questions Claude can answer with evidence. For example: "Are they hiring salespeople?", "Who owns outbound today?", "Did they raise in the last year?"

Keep the list short, make every question answerable from public information, and ask for a verdict plus the evidence. The answers do double duty: they decide who gets contacted, and they become the specifics your first line refers to.

2. Give Claude data tools over MCP

MCP (Model Context Protocol) is the open standard that lets Claude call tools. Many data providers now ship MCP servers, Apollo and Clay among them. In Claude Code you add a server with claude mcp add; in Claude Desktop and Cowork many tools install as connectors from settings.

Connect one data source to start. More tools is not better: every server adds tool definitions to Claude’s context, and a crowded context makes worse decisions. See MCP for sales teams for what to connect.

3. Put the work somewhere that remembers

Chat history is not a system of record. Start a new session and yesterday’s research is gone, so Claude researches the same accounts again. You have three options:

• A spreadsheet. Easy to start, but Claude has to re-read it every session, and there is no place for evidence, sources or how old an answer is.
• Your CRM. Good for records, but follow-ups become tasks rather than sequences, and its fields were not designed for research with evidence.
• An MCP harness such as Harp, built for an agent: accounts, answers saved with evidence and age, sequences, and a short tool list that keeps whole account lists inside Claude’s context.

The test is simple: can tomorrow’s session see what today’s session learned, without reading everything again?

4. Send from your own mailbox, slowly

A connector can send an email the moment Claude asks. That is the problem: outbound needs pacing. The basics:

• Send cold outreach from a separate domain, not your main one, and give new mailboxes a few weeks to warm up.
• Use a real Google Workspace mailbox rather than shared sending servers.
• Space sends minutes apart and cap each mailbox per day.
• Land emails in the recipient’s working hours.
• Stop a person’s sequence the moment they reply, bounce or unsubscribe.

5. Decide who answers replies, and how fast

The first reply is the moment that matters, and a reply that waits until your next session is often a lost meeting. Either you watch the inbox, or the reply wakes your agent.

If the agent answers, give it a short list of product facts it may state, and tell it what goes to a human: pricing, security questions, anything it cannot answer from the facts. Harp’s reply handling works this way: a reply stops the sequence, wakes your agent within minutes during your working hours, and the agent answers the email or books the meeting. See Reply handling.

6. Add LinkedIn carefully

LinkedIn adds a second channel and real risk. Keep volumes human, warm new accounts up, browse by clicking rather than jumping between profile URLs, and confirm you are on the right person before anything is sent. How to automate LinkedIn without putting your account at risk goes through each practice.

7. Run a daily loop

Once the pieces are in place, the day is a few prompts:

• "Start with a briefing: what is due today, what is held, what research is stale?"
• "Fill the research gaps for the next 20 accounts and save the answers with evidence."
• "Enroll the qualified contacts in the re-engagement playbook."
• "Review the drafts going out tomorrow and tighten the first lines."

Where Harp fits

Harp is the MCP harness for pipeline generation: the memory, sequences, sending, reply handling and LinkedIn safety around the agent you already use. You bring your own AI plan; Harp sells no AI credits. If you run this setup by hand today, Harp vs DIY outbound with Claude shows where the two differ.