Using AI agents to manage LinkedIn replies
Connect Claude or ChatGPT to your LinkedIn sales pipeline using MCP and APIs to draft personalized replies and manage leads without copy-pasting.
Get startedShort answer
You can connect Claude or ChatGPT to your LinkedIn sales process through a Model Context Protocol server or a REST API. The agent then reads your lead profiles, thread history and replies directly, so it drafts answers with real context instead of guessing from one pasted message.
The problem with copy-pasting
Most people copy a LinkedIn message, paste it into ChatGPT, and ask for a reply. That works for three conversations. At twenty, you spend more time moving text between tabs than selling.
The bigger issue is context. A pasted message carries no job title, no company size, no memory of what you said last week. The model fills the gaps with guesses, and guesses read like a template.
There is also the sorting problem. Replies arrive mixed together: a yes, a not-now, a wrong person, an annoyed one. If you draft each one in the order it landed, you answer the least important one first.
Copy and paste
- Copy the message into a chat window
- Ask the AI to make it sound professional
- Paste back a reply that says "I would love to learn more"
- Forget which lead is at which stage by Thursday
Agent with context
- The agent reads the profile, the thread and your notes
- It names a detail from the lead's own post
- It drafts against your playbook, not a generic tone
- The draft sits next to the lead it belongs to
How MCP connects AI to your leads
Model Context Protocol is an open standard. It lets AI clients such as Claude Desktop, ChatGPT desktop or Cursor call tools on a server you control, so the model fetches data instead of waiting for you to paste it.
The server exposes a small set of actions: list new replies, get a lead's profile, get the thread history, save a draft. The model picks which one to call based on what you asked for.
That is the difference between a chatbot and an agent. A chatbot answers what you typed. An agent goes and looks something up first, then answers.
- New replies since a timestamp, with the lead attached
- The lead's headline, company and recent posts
- The full thread, so the model knows what was already said
- Your playbook file and calendar link
Setting up your AI sales workflow
You need three things: a client that speaks MCP, a server that can see your LinkedIn data, and a written version of how you sell. Set them up in this order.
- 1
Pick the client you already use
Claude Desktop, Claude Code, Cursor and ChatGPT all accept MCP servers now. Use whichever you keep open all day. Adding a server is a settings screen, not a project.
- 2
Connect the server with OAuth or an API key
Chat apps usually take an OAuth flow: click connect, approve, done. Scripts and scheduled jobs take an API key. Either way the credential belongs to your sales tool, not to the model.
- 3
Write the playbook as a file
One page. What you sell, who it fits, the three objections you hear most, your calendar link, and the phrases you never want to see. Keep it in a file the server can read so every draft uses the same version.
- 4
Run one loop by hand first
Ask the agent for new replies, read what comes back, ask it to draft one, and check the draft. Do that for a week before anything sends on its own.
Writing a prompt for sales drafts
The reply is only as good as the constraints. Tell an AI to be professional and it writes "I hope this message finds you well" and "I wanted to reach out". Prospects read that as a mail merge.
Give it a role, a length limit, and one rule that forces it to use real detail from the lead. Then list the phrases you never want to see, so you stop editing the same sentence every time.
Example
- Suggested reply
Hi Sarah, your post about the hybrid rota at Acme stuck with me. The scheduling part is usually where it breaks, and I've watched three agencies your size work through it. Worth 10 minutes?
- Prospect
Honestly the tooling is the bigger headache for us. Can I see a demo?
- You
Yes. Here's my calendar, take whatever slot suits.
Paste this job
System prompt: you write short LinkedIn replies for a B2B seller. Goal: book a 15 minute call, or get a clear no. Rules: 1. Three sentences maximum. 2. Open with one specific detail from the lead's profile or their last message. 3. Never write "I hope this message finds you well", "I wanted to reach out", "circle back" or "touch base". 4. One question at the end, and it must be easy to answer. 5. Write like a person typing on a phone. No greeting line, no sign-off. 6. If the lead said no, close the loop in one line and stop.
Sorting replies by intent before you draft
Twenty replies are twenty different jobs. One is a yes, one is a no, one is a referral to a colleague, one is a question about price. Drafting in arrival order means the urgent ones wait behind the dead ones.
Ask the agent to classify before it writes. A short label per reply is enough: interested, question, not now, wrong person, not interested. Then draft only the first two.
This also protects you. A reply that says "we already use someone" should get a short, polite close and a note to check back in six months, not a cheerful pitch.
Interested
- Intent
- They asked a question or named a time
- What it usually means
- Answer today, with a calendar link
Question
- Intent
- Price, scope, or how it works
- What it usually means
- Answer plainly, then ask for the call
Not now
- Intent
- Timing, budget, headcount
- What it usually means
- Agree, ask when to check back, set a reminder
Wrong person
- Intent
- They named a colleague or forwarded you
- What it usually means
- Ask for the intro, thank them once
Not interested
- Intent
- A clear no
- What it usually means
- Close the loop in one line, no pitch
Managing the reply loop
Drafting is the easy half. The hard half is deciding what the agent is allowed to do without you.
Pick a mode per campaign and write it down. Hand off at the first reply means the AI stops and you take over. AI replies until interest means it handles the small talk and hands over when someone asks about price or a call. AI replies until a meeting is booked means it runs the whole thing with your calendar link.
Whatever you pick, read the first fifty drafts. You are looking for one thing: whether the model invents details. A draft that says "I saw you worked at Northwind" when they didn't is worse than a boring one.
Scan
- Stage
- Pulls new replies and labels intent
- What the agent does
- Skim the labels, open the interested ones
Draft
- Stage
- Writes against the playbook and the thread
- What the agent does
- Read for invented facts, not for style
Send
- Stage
- Sends in hand-off mode or on its own, per campaign
- What the agent does
- Approve in hand-off mode, spot-check otherwise
Follow up
- Stage
- Waits, then nudges once with new information
- What the agent does
- Kill the sequence the moment someone says stop
Doing this with Kinlead
Kinlead runs a REST Agent API and an MCP server, so Claude, ChatGPT, Cursor, Claude Code and Codex can find leads, check replies and draft answers. Chat apps connect with OAuth, scripts use API keys.
Sending stays inside Kinlead. Outreach goes out from your own LinkedIn account through Unipile, at most one action every five minutes, inside the daily limits and send window you set. Replies from every campaign land in one inbox, sorted by intent, and you get an email when someone replies.
Each campaign also decides how replies are handled: hand off at the first reply, let the AI reply until there is real interest, or let it reply until a meeting is booked with your scheduling link. Pro is $49 a month after a 7-day trial.
Read the first fifty drafts before you let any campaign send on its own. You are checking for invented facts, not for style.
Frequently asked questions
Related
Get started with Kinlead
Prefer markdown? Using AI agents to manage LinkedIn replies.md