Outbound Agent It finds clients and emails them for you.

A real fact beats the template

In one of my agency's campaigns, emails written around a real fact about each company got 6 replies out of 66. The same offer as a template got 0 out of 162. Outbound Agent writes the first kind, at volume, and sends nothing you haven't seen until it has earned it.

Each one does a different job. The last one is the best.

They run as one pipeline, and each one also works on its own: ask Claude in plain words (“check my domains”, “answer my replies”) or type its name, like /outbound-domains.

1. 1
   outbound-icpWho to target

   Sets everything up, reads your website and your best customers, and writes your offer card and the exact people to target.

2. 2
   outbound-leadsThe list

   Only decision makers, only verified emails, and a registry so nobody is ever contacted twice.

3. 3
   outbound-researchOne real fact

   One true, recent fact per company, with its source and its date. No fact, no email.

4. 4
   outbound-copyThe email

   A short email built around that fact, in plain words, with one easy question.

5. 5
   outbound-fact-checkProve it wrong

   A second reviewer with fresh eyes checks every sentence against the source: same tense, who did what, nothing added.

6. 6
   outbound-follow-upOne more, with something new

   One follow-up in the same thread that adds something new. It stops the moment they reply.

7. 7
   outbound-domainsStay out of spam

   Every morning: SPF, DKIM, DMARC, MX, blocklists and the mailbox ramp, with real DNS checks.

8. 8
   outbound-sendApprove, then load

   You approve your first 5 emails one by one. Then it loads your sender with daily caps and proves it loaded.

9. 9
   outbound-repliesAnswers in minutes

   Sorts every reply and answers with two free slots from your calendar. A second pass scores it, and under 80 it rewrites.

10. 10
    outbound-bookingThe invite

    Re-checks the slot, sends the calendar invite with a Meet link, and confirms in the thread.

11. 11
    outbound-linkedinSame lead, on LinkedIn

    Connection note and DM for the same leads, from the same fact. Drafts that you send.

12. 12
    outbound-recapAfter the call

    The recap email from the call transcript, with clear next steps. Saved as a draft, never sent.

13. 13
    outbound-learnCopies what worked

    Every evening it reads which emails got replies and writes tomorrow's batch to copy what worked, with strict evidence rules.

    The best one

One link, or two commands

From zero to your first batch, step by step

About 30 minutes the first time, then a few minutes a day. Everything lives in plain files in ~/outbound-agent/ on your computer.

Nothing is sent until you approve 5 emails

• An approval belongs to the exact text. If the email changes afterwards, it needs approving again. A “looks good” in chat doesn't count.
• Improve becomes a rule. Your comment is saved and every future batch follows it. The most recent comment wins.
• After 5 approvals the campaign runs on its own and keeps learning from your comments and from the replies.
• A new offer, language or audience is a new campaign: the count starts again from zero.
• The same goes for answers to replies: drafts until you've approved 5 and switched them on.

Every evening, it copies what worked

outbound-learn reads which emails got replies, by angle (what the fact was about: a new plant, funding, a new leader…) and by version of each part of the email. Then it decides what tomorrow's batch should copy, and writes it, sized on the real free slots of your mailboxes. So you stop guessing what to send, and it still sounds like you. It decides with evidence rules, never on two replies:

• Under 5 replies in total, there's no winner, and it says so: tomorrow follows the quality rules and keeps rotating.
• If someone replies twice, the best reply counts: a “Thanks!” after a “yes” doesn't erase the yes.
• Versions only gain weight after 40 sends, every live version keeps at least 15% (it keeps testing), and a version with 60 sends and zero replies is retired and replaced.
• One new hypothesis per batch, written with a date to check it. And it never writes batches days ahead: they'd miss what it learned in between.

The domain check, on a real domain

This is outbound-domains run on growthcab.com, my agency's main domain, on 19 Oct 2026. Copied from the terminal as it came out.

YELLOW growthcab.com
   ! Spamhaus DBL not verifiable (127.255.255.254 query refused (public resolver))
   - SPF: ok, ends with ~all, 1 lookups
   - DKIM: found on google ~2048-bit
   - DMARC: p=quarantine; reports go to the rua address
   - MX: 1 aspmx.l.google.com., 10 alt3.aspmx.l.google.com., 10 alt4.aspmx.l.google.com., 5 alt1.aspmx.l.google.com., 5 alt2.aspmx.l.google.com.
   - age: 1211 days, website: 200 https://www.growthcab.com/

RESULT 0 red, 1 yellow, 0 green  (2026-10-19T11:49:05-03:00)

Why yellow: a blocklist that can't be verified counts as unknown, never as clean. Spamhaus refuses queries that come through public DNS resolvers, so without a key it's “not verifiable”: a free Spamhaus Data Query Service key fixes that. SURBL counts as clean only if its own test domain shows up as listed on the same resolver, in the same minute.

Honest limits

• It doesn't buy domains, create mailboxes or change your DNS. It tells you exactly what to fix.
• It doesn't guarantee inbox placement or replies. Your list, your offer and your domains decide that.
• It sends nothing before your first 5 approvals, and answers to replies stay drafts until you switch them on.
• It never automates LinkedIn: notes and DMs are drafts you send yourself.
• It doesn't store API keys or install background services. Routines run only if you create them.
• It isn't legal advice. Cold email rules depend on the recipient's country, for example CAN-SPAM in the US and consent rules in the EU and UK. Check yours.
• The domain check needs dig: macOS and Linux have it, on Windows use WSL.

Short answers