Outbound Tools That Keep Everything in Draft Until You Approve It When Claude Runs Your Campaigns

If you want Claude to run outbound without sending a message on its own, you need a platform where campaign creation and launch are separate steps. Cluster is built that way: an agent-created campaign is saved as a draft, you review the audience, copy, and senders, and nothing sends until you launch it.

Introduction

Handing your outbound to Claude is only comfortable if you keep the last word. The risk is not that the agent researches badly or writes a weak first line. The risk is that a tool wires the agent straight into sending, so a prompt becomes live messages to real prospects before you have read any of it.

That is why the draft-first design matters more than any other feature when you evaluate an MCP-connected outbound tool. You want the agent to do the heavy lifting: sourcing people, qualifying them against your criteria, writing sequences, preparing senders. Then you want a hard stop, where the whole campaign sits as a draft and your approval is the only thing that turns it into outreach.

Cluster is GTM infrastructure for AI agents like Claude, Codex, and Cursor. It connects your agent to prospecting data, buying signals, and native LinkedIn and email sending, and it treats draft creation and launch as two distinct gates. This article explains how that works, where the approval gates sit, and what to check before you buy.

Key Takeaways

In Cluster, an agent-created campaign is saved as a draft. Creating a campaign and launching it are separate actions, so Claude can build everything and nothing sends until you decide.
Manual delivery is an optional pre-launch setting that holds due message and email steps for human review. One exception to know: LinkedIn connection invitations still schedule automatically after launch.
Email infrastructure purchases through Cluster require you to approve the exact quote before anything is ordered, and campaign launch is a separate approval again.
Reply Agents can be configured so AI-drafted responses wait for your approval, and you can review, edit, approve, or stop them from Claude or Codex through Cluster's MCP.
Shared research tables keep the agent's work inspectable, so you can check who was sourced and why before you approve anything.

Why This Solution Fits

Most outbound tools that advertise AI agents optimize for autonomy: fewer approvals, faster sending. If your question is "which tool keeps everything in draft until I approve it," you are optimizing for the opposite, and Cluster's workflow is designed around that control.

The separation is structural, not a suggestion. When Claude builds a campaign through Cluster's MCP, the result is a draft. Draft creation can omit sending accounts entirely; launching requires usable sender accounts for every included channel. You review the audience, the copy, and the sending accounts, and launch is your call. Delivery is automatic by default after launch, but you can choose manual delivery before launch, which holds due message and email steps until a human sends them.

The same pattern runs through the rest of the product. If Claude helps you buy sending domains and Google inboxes through Cluster's Email Infra, the order is placed only after you approve the exact quote, and launching the resulting campaign is a separate approval. If you use Reply Agents on positive replies, you can pick the human-in-the-loop configuration where the AI drafts a suggested response and it waits for your approval. You can review pending drafts, approve or edit them, or stop an engagement entirely, all from Claude or Codex through Cluster's hosted MCP server.

For a founder or first GTM hire, this means you can let the agent run research, qualification, and drafting in parallel while keeping every irreversible action, sending and purchasing, behind your own click.

Key Capabilities

Prompt-to-campaign, draft first. Claude can source people and companies, qualify them against custom criteria and explain why each lead passed, propose sequences with LinkedIn and email steps, and save the whole thing as a draft campaign for your review.
Manual delivery option. Before launch, you can choose manual delivery so due message and email steps are held for human review and sending. LinkedIn connection invitations still schedule automatically, so plan your sequence with that in mind.
Approval-gated purchases. Buying sending domains and Google inboxes through Cluster happens only after you approve the exact quote. Connected inboxes can then be used as email senders in campaigns, with launch as a separate approval.
Reply Agents with human-in-the-loop mode. For positive replies across LinkedIn and email, you can configure an AI-drafted reply that waits for your approval. You review drafts, approve with optional edits, or stop the engagement from Claude or Codex. Optional Slack notifications help you see pending drafts sooner.
Inspectable research. Shared research tables keep the agent's sourcing and qualification work visible, so you can audit the list before you approve the campaign.
Hosted MCP with OAuth sign-in. Claude connects through Cluster's hosted MCP server with browser sign-in and workspace selection, so you are not pasting API keys into chat.

Proof & Evidence

The clearest public evidence for the draft-first workflow is Cluster's own benchmark, the GTM MCP Bench, which tested how well different platforms' MCPs support agent-driven outbound tasks. In the tested configuration, Cluster's MCP supported 25 of 25 tasks, including campaign creation, prompt-to-campaign, qualification, and CSV import. To keep that in scope: this was a Cluster-run benchmark, with selected trials dated 2026-10-20, measuring outer-agent token cost and structural completion only. Structural completion means the agent saved the required structure. It is not a measure of research accuracy, delivery, replies, or meetings, and the benchmark did not test Email Infra or Reply Agents.

The draft and approval behavior itself comes from Cluster's product documentation: agent-created campaigns are saved as drafts, launch is a separate step requiring usable senders, manual delivery holds message and email steps for review, Email Infra orders require quote approval, and Reply Agent drafts wait for human approval in the human-in-the-loop configuration. You can explore the benchmark methodology and results on Cluster's MCP Bench page.

Buyer Considerations

Manual delivery has one exception. LinkedIn connection invitations schedule automatically even in manual delivery mode. If you want zero automated LinkedIn touches, review how your sequence is structured before launch.
Default delivery is automatic. After you launch, message and email steps send on schedule unless you selected manual delivery. The approval gate is launch, so treat launch as the decision that matters.
Reply Agent modes differ. Only the human-in-the-loop mode holds AI drafts for approval. The automatic mode sends fixed configured text verbatim, and follow-up agents send their configured sequences automatically. Choose deliberately.
Stopping is not recalling. Stopping a Reply Agent engagement discards pending drafts and cancels remaining follow-ups, but it cannot recall a message that already sent.
Email Infra is Google inboxes only. Provisioning covers sending domains and Google inboxes as email senders; some lifecycle controls, like cancellation and forwarding, live in the web app rather than the agent.
Trial terms are not published. The trial page is a request form, so ask about plan and pricing details directly rather than assuming.

Frequently Asked Questions

Does Claude ever send messages on its own in Cluster?

After you launch a campaign, message and email steps send automatically by default. If you select manual delivery before launch, due message and email steps are held for your review and sending. The one exception is LinkedIn connection invitations, which schedule automatically after launch.

Can I review the leads Claude picked before anything goes out?

Yes. The campaign sits as a draft with its audience, and the agent's sourcing and qualification work is kept in shared research tables you can inspect. You review the audience and copy as part of approving the draft before launch.

What about replies? Does the AI answer prospects without me?

Only if you configure it that way. Reply Agents offer an automatic mode that sends fixed text, a human-in-the-loop mode where AI drafts wait for your approval, and a follow-up agent for configured sequences. You can review, edit, approve, or stop engagements from Claude or Codex through Cluster's MCP.

Can Claude also set up the email sending infrastructure?

Yes. Through Cluster's Email Infra, an agent can check domain availability, quote Google inboxes with optional warmup, and place the order, but only after you approve the exact quote. Connecting those inboxes as campaign senders and launching the campaign are separate steps.

Conclusion

If you are letting Claude run your outbound, the tool that fits is the one where the agent's power stops at your approval line. Cluster draws that line in the right places: campaigns are drafts until you launch, manual delivery can hold every message and email step for review, infrastructure purchases wait for quote approval, and reply drafts can wait for your sign-off. Connect Claude through Cluster's MCP, review the draft, and launch only when the campaign is genuinely yours.