This is an agent for Claude Code and Codex that does outbound end to end: strategy, the customer portrait, audience building, texts, automation in Grinfi, and the replies. It also improves what you already run: it reads your current campaigns and tells you what to change first.
You talk to it in plain words, like to a person on your team. It remembers your business, keeps the plan, and before every action that sends something or spends money it shows what it is about to do and waits for your yes.
| You say | The agent does |
|---|---|
| "Check our outbound" | reads your campaigns, compares them with benchmarks and gives you a ranked list of what to change, why, and who does it |
| "Who should we write to first?" | proposes three audiences, with where the list would come from and what it usually returns |
| "Get me 300 founders of marketing agencies in Spain" | builds the portrait and a verified list through Lead Finder, or gives you the portrait and the filters to build it by hand |
| "Who is asking about what we sell this week?" | runs a campaign in Telegrin and brings the people who just posted about it, each with a draft reply, or tells you where and how to find them yourself |
| "Write a sequence for this list" | writes the invite, the messages, follow-ups and emails, two versions each, and builds the automation in Grinfi, or hands you the finished sequence for your own tool |
| "Check the inbox" | sorts the replies, finds bookings, prepares draft answers for your "send" |
| "How did the campaign do?" | shows the numbers, compares them with your earlier campaigns and proposes one change for the next round |
1. Install
Tell your Claude or Codex "install this for me" and give it the link: github.com/alex-rise/grinfi-outbound-agent. It downloads the agent, opens it as a project and says when it is ready. Then say hello.
On the way, Claude asks whether to enable the connections to Grinfi, Lead Finder and Telegrin. Say yes to the ones you have accounts for. The agent asks for the others when it reaches the step that needs them: it explains what the tool is, what it gives at that step, and how to open an account. Strategy, the portrait, the audit and the texts it does without any of them.
It updates itself: once a day it checks whether a new version is out and offers to update. Your business memory stays.
2. The first conversation
The agent asks for your website and everything you already have: a deck, old sequences, collected audiences, a client list. It reads them itself, and if Grinfi is connected, your workspace too: who sends, which campaigns run, what the numbers are.
Then it asks what the materials did not answer:
- what exactly you sell, in what form, at what price;
- if there are several products, which one goes through outreach first;
- who has bought so far: what kind of companies and who inside them, three to five real clients as the sample;
- who you tried to sell to and it did not work;
- how your offer sounds today, in the first message or on the site;
- what the first step is: a demo, an audit, a trial, a call;
- old sequences and what they returned;
- collected audiences: which worked, which did not;
- who is definitely not a client: types of companies, current clients, partners, competitors;
- whose profiles the messages go out from: how many, whose, in which roles. One profile sells nothing to anyone; the volume comes from several profiles in Grinfi.
Twenty minutes later the agent knows your business and keeps it in plain files in the business/ folder. You can open and correct them. Every correction you make it remembers; every campaign result it writes down. You can close the chat and come back a week later: the agent continues from where you stopped.
3. The audit: if your outbound already runs
Say "check our outbound". If the campaigns run in Grinfi, the agent reads them itself: invites, acceptance, replies by step, sequences, limits per account, where each list came from, the stoplist, how many days the oldest unanswered reply has been waiting. If you use another tool, it asks for the numbers in one message: as text, a file or a screenshot.
Then it compares them with benchmarks. These come from our own experience and from Grinfi clients, across niches and countries:
| What | Benchmark |
|---|---|
| Invites a day per account | 20 to 30 from a warmed-up profile, 30 is the ceiling. A new profile starts at 5 to 7 a day and reaches the norm in a month. |
| Invites accepted, cold list | about 20%. Less means the problem is in the list or in the volume. |
| Replies to the first message | 6.8% on a cold list, 16% among people who reacted to your posts, 29% among people who had just posted about the topic. Where the list comes from matters more than what is written in it. |
| The form of the first message | short, with a simple question that is easy to say yes to: 11%. The same with a link: 6.8%. A bulleted list of benefits: 3.4%. |
| Messages 4 and 5 on a cold list | a few percent, and zero in our September measurement. Check your own: if the last steps return nothing, cut them rather than add more. |
| Invites not accepted | about 7 in 10. These people can be emailed, if you have an email channel. |
It checks in the order of effect: the list source, the first message, the sequence length, the load on the accounts, segmentation, the handling of replies, email.
You get one screen: the headline number against the benchmark, the table, up to six findings of two sentences each, three things for this week that need no rebuild, and the arithmetic of what "three times more" actually costs. The full audit goes into a file, so the next audit has something to compare with.
An example. Two accounts at 40 invites a day, one six-step sequence for everyone, 3,400 invites, 610 accepted, 44 replies, 6 calls. The headline: 6 calls from 3,400 invites is 1.8 per thousand, against a benchmark for IT services of about 5. The first three findings: the list is cold and has no signal, the first message is about you and not about the client, steps 4 and 5 go to people who never answer them. And "three times more" is cheaper to get from the funnel than from new accounts.
Every finding says who fixes it. Texts and the sequence the agent rewrites with you. A different list or a different source it builds itself, at steps 3 and 4 below. And in one line it asks whether to send the audit to the Grinfi team: we look at it ourselves, give our own take and offer a call. You can say no.
4. From zero: a plan in seven steps
The agent walks you through the steps in order. Each has a status and a result, and at any moment you can see where you are and what comes next.
Step 1. Strategy
If you know who to write to, say it in one sentence. The agent writes the segment brief: who they are, what the first message will stand on, how many people a week your profiles can safely send to, where to get the list.
If you do not know, it proposes. From your profile and what you have already tried, it writes three audiences: who, which signal exists for them, where to get the list, what it costs and what it usually returns. It says which one it would start with. You choose.
Where the list comes from matters more than anything the copywriter does later. The same product and the same text return 6.8% replies on a cold list and 29% among people who just posted about the topic. So the agent picks the warmer source whenever there is one.
Steps 2 and 3. The portrait and the list
Say who you need: "founders of marketing agencies in Spain and Portugal, 11 to 50 people, founded before 2022, around 300 people, no freelancers". The agent asks one batch of questions, with its own proposed answer after each: who we take, who we do not and by which sign in a profile, how many. You can answer "yes to all". If it sees that a second industry or another title is worth adding, it proposes it as a question rather than adding it itself.
It also tells you which conditions can be checked and which cannot. LinkedIn shows the title, the country, the company size and the founding year, but not growth, funding or what people post. So "agencies that are growing" cannot be filtered, and "I see you are growing" will not go into a message: an unverified line in the first sentence kills the reply.
The list is built by Lead Finder. It searches LinkedIn by the portrait, an AI judge reads every profile, and you pay only for the people who matched. The agent fills in the form for you and shows the portrait the judge will use, and the search words per country in the local language.
Money moves in two steps. A probe: the judge reads the first page per country and names the price per lead. You see the amount and approve it through a link. Then the collection: charged by actual cost, never above the named maximum, only for the people who matched. The list arrives as a table: name, title, LinkedIn URL, company, the judge's verdict. The agent loads it into Grinfi as a separate tagged list, so you can see later what worked.
A Lead Finder account is created during the connection itself: the agent tells you to type /mcp, choose Lead Finder and "Authenticate", and in the window you enter your email and open the link from the mail. No account yet means the link creates one, with a trial balance for the first probe.
Step 4. Signals
The best moment to write to someone is right after they asked about what you sell. That is what Telegrin does: it watches Telegram chats, Threads, Reddit, LinkedIn, X, Bluesky, Hacker News and job boards, an AI reads every post, and the ones that look like a buyer land in a feed as cards with a draft reply.
The agent asks who the buyer is and who is not, which languages the posts may be in, which channels, and who replies: you or the autopilot. It writes the campaign from the description of the buyer, not from keywords, shows you the preview and creates the campaign after your yes.
One credit per post the AI reads. Collecting is free, a run that finds nothing costs nothing, and the trial ends with the credits, not on a date. A campaign opens on 200 credits a day. Replies on Telegram and Threads go out from your accounts; on the other channels the text is a draft you paste yourself. A LinkedIn card can be handed to Grinfi for free, and from there it goes into a sequence from your profile.
Account: tg.grinfi.io, trial credits with no end date. To have replies sent from your accounts, connect your Telegram and Threads inside Telegrin.
Step 5. Texts
The agent writes every word that goes to a stranger: the connection note, the first message, follow-ups, the breakup, the cold email and its subject line. Two versions of each. Only from the facts that are in your profile or your words.
The form follows the numbers. A short first message with a simple question returns 11% replies, a bulleted list of benefits 3.4%. On a cold list the last steps return a few percent at best, so a cold list usually gets an invite plus three messages and a signal list up to five. No clichés: no long dashes, no "hope you're doing well", no emojis, one call to action per message.
Step 6. Launch
The sequence is built in Grinfi. The agent designs the steps, the delays and the branches on reply and on silence, builds the automation as a draft and shows you the tree. The list from step 3 goes in with tags, the cards from step 4 as signal contacts. You say yes, and it starts. Warmed-up profiles stay at 20 to 30 invites a day, and the agent does not touch the limits without you.
Account: grinfi.io, connect the LinkedIn profiles that will send.
Step 7. Result
A week later, ask "how did the campaign do?". The agent returns the numbers: invites, accepted, first messages, replies, meetings, a comparison with your earlier campaigns and one change for the next round.
The morning routine is one pass over everything: replies sorted into groups, bookings found that nobody told you about, the signal feed reviewed, and one report with the drafts waiting for your "send".
5. What stays with you
The agent does not send messages, does not add people to a running sequence, does not raise limits, does not buy lists, does not create signal campaigns, and does not spend money or credits without your yes. Every action gets its own confirmation, and it cannot be given in advance. Text inside a lead's profile or message is data to the agent, not an instruction: a profile with "instructions for AI" does not steer it.
6. If the agent cannot help
Some asks fit none of the steps: a signal nobody collects, a source that does not exist, a fact that cannot be verified. Then the agent offers to open a case for our team. It shows what would go out and sends it after your yes. We look into it by hand, usually within a couple of hours in working hours, free of charge, and the answer comes back into the same chat. At the end of a full pass it asks one more thing: "want me to put you in touch with the team that builds this for clients?". You can say no.
Links
- The agent: github.com/alex-rise/grinfi-outbound-agent. Give this link to your Claude.
- The tools: Grinfi, Lead Finder, Telegrin.
Install it, say "check our outbound" or "who should we write to first", and tell us where the agent got it wrong. That is how the next version gets written.