Turn Your AI Agent Into an AI SDR: LinkedIn Outreach via the proReach MCP Server

Last week we watched an AI agent take the prompt “I run a small SaaS for Danish dentists and I need customers” and, in seven conversational turns, do the work a junior SDR would spend an afternoon on: it checked which LinkedIn account was connected, picked a cold-outreach sequence, wrote a connection note in Danish appropriate for clinic owners, added the prospects, ran a pre-launch safety check. Then it stopped, reported exactly what would happen next, and asked for permission before sending anything. Nobody clicked through a dashboard. The entire thing happened inside a chat.

That is an AI SDR: an AI sales agent that handles the top of your pipeline the way a sales development rep would. The test above was the acceptance criteria for something we just shipped: proReach now has an MCP server, which means any MCP-capable AI assistant (Claude Code, Cursor, Windsurf, Claude Desktop) can operate your LinkedIn outreach for you. This post explains what that actually means, how to set it up in about five minutes, and the guardrails that make “let an AI run my outreach” a sane sentence.

What “MCP” means if you don’t follow AI infrastructure news

MCP (the Model Context Protocol) is an open standard, originally from Anthropic, that gives AI assistants a way to use outside tools. Think of it as a plug socket: a product exposes its features as tools, your AI app connects to that socket, and from then on the assistant can call those tools mid-conversation the same way it can search the web or read a file.

proReach’s socket exposes eight tools that cover the entire customer-acquisition loop: checking your workspace status, listing campaign sequence templates, creating a campaign with personalized copy, adding leads, launching (gated; more below), pausing, reading campaign metrics, and summarizing which prospects have replied and are waiting on you. Every tool answers with plain-English next steps, so the agent always knows what to do next without you spelling it out.

Setup: five minutes, three steps

1. Get a key. In proReach, open Settings → AI agents (MCP) and create an API key. It is shown once: treat it like a password, and revoke it from the same screen if it ever leaks.

2. Connect your AI app. For Claude Code, it is one command:

claude mcp add --transport http proreach https://proreach.ai/api/mcp \
  --header "Authorization: Bearer prk_YOUR_KEY"

Cursor users add the same URL and header to ~/.cursor/mcp.json; Claude Desktop connects through the mcp-remote bridge. The exact config blocks for each app are in the key-creation screen, so you can copy them with your real key already filled in.

3. Optional but recommended: install the skill. We published a companion playbook on PolySkill that upgrades your agent from “can operate the tools” to “thinks like an outreach consultant.” It makes the agent qualify your offer first (what do you sell, who buys it, what outcome do they get) before it writes a single line of copy:

npm install -g @polyskill/cli
polyskill install @mrspacemann/proreach-outreach

Then just talk to your agent: “I need customers, set up LinkedIn outreach for me with proReach.”

What the agent actually does (and what it refuses to do)

The interesting part of this design is not what the AI can do. It is where the hard lines are. We drew three of them deliberately:

The AI never touches LinkedIn.Agent tools only read and write campaign plans inside proReach. The actual LinkedIn activity (connection requests, messages, profile visits) is performed by proReach’s own automation on its own schedule, inside LinkedIn-safe daily limits, working hours, account warm-up, and automatic cooldowns. An over-eager agent physically cannot burn your account by spamming, because there is no tool that sends anything directly.

The AI never sees your LinkedIn credentials. Connecting your LinkedIn account happens in the proReach web app, through the same secure flow as always. There is no MCP tool that accepts a password. If an agent (or a manipulated prompt) tries to collect one, there is simply nowhere to put it.

Launching requires your explicit yes.The launch tool is two-step by design: called once, it returns a preflight report showing how many leads will enter the sequence, from which account, and any blockers like an expired session or a plan limit. The agent has to show you that report and get your go-ahead before it can confirm. In our seven-turn test above, the agent hit a plan limit at preflight, explained it, and explicitly declined to pause the existing live campaign because its instructions say never to touch campaigns the user didn’t ask about. That is the behavior you want from something acting on your behalf.

The AI SDR category, done the boring way

There is a growing category of AI SDR products that promise a robot salesperson end to end. Our view is more boring and, we think, more correct: the strategy layer (who to target, what to say, how to interpret results) belongs in a conversation with an AI that knows your business. The execution layer (real LinkedIn accounts, proxies, sending caps, warm-up, reply detection) belongs in infrastructure that has been hardened against LinkedIn’s limits for a long time. The MCP server is simply the clean seam between the two.

Concretely, that means your agent is genuinely good at the parts language models are good at: it will push back on a fuzzy offer before writing copy, produce a connection note in your prospect’s language, keep the first touch claim-free (asking beats selling in connection requests), and translate acceptance and reply rates into plain-English decisions: tighten the audience, personalize harder, or scale up. And when replies arrive, the agent’s job flips to making sure you answer them: replies are where customers come from, and the inbox summary tool exists mostly to nag you about that fact.

The honest limitations

Two things the agent will hand back to you today. Lead discovery by search (pasting a LinkedIn or Sales Navigator search and having the list flow in) happens in the proReach web importer, not in chat yet; the agent adds prospects from profile URLs and deep-links you to the importer for bulk discovery. And replying to prospects happens in the proReach inbox: the agent reads and drafts, but the send button is yours. Both boundaries are deliberate for now, and both are on the roadmap to soften.

Try it

If you already use proReach: create a key in Settings → AI agents (MCP), run the one-liner above, and ask your agent to check your workspace. If you don’t: start a free trial, connect your LinkedIn account, and let your agent take it from there. The first thing it will do is tell you exactly what is missing and where to click.

Frequently asked questions

Related reading

• proReach features The full outreach and content platform the MCP server drives.
• How to Write a LinkedIn Post That Actually Gets Engagement The same first-line discipline that makes posts work makes connection notes work.
• Best Time to Post on LinkedIn Timing logic your agent inherits automatically through working-hours sending.